Merge branch 'develop' into sphinx_doc

library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <array>
+#include <algorithm>
+#include <thread>
+
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct ReferenceBatchNormFwd : public device::DeviceBatchNormFwd<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                 const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* bnScale,
+                 const BiasDataType* bnBias,
+                 double epsilon,
+                 const YElementwiseOp y_elementwise_op,
+                 YDataType* p_y,
+                 MeanVarDataType* resultSaveMean,
+                 MeanVarDataType* resultSaveInvVariance,
+                 double averageFactor,
+                 MeanVarDataType* resultRunningMean,
+                 MeanVarDataType* resultRunningVariance)
+            : reduceDims_(reduceDims),
+              bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              bnScale_(bnScale),
+              bnBias_(bnBias),
+              y_elementwise_op_(y_elementwise_op),
+              p_y_(p_y),
+              resultSaveMean_(resultSaveMean),
+              resultSaveInvVariance_(resultSaveInvVariance),
+              resultRunningMean_(resultRunningMean),
+              resultRunningVariance_(resultRunningVariance)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            // get invariant_dims[] and invariant_lengths[]
+            for(int dim = 0, i = 0; dim < Rank; dim++)
+                if(std::none_of(
+                       reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                {
+                    invariantDims_[i]     = dim;
+                    invariant_lengths_[i] = xyLengths[dim];
+                    i++;
+                };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = xyLengths[dim];
+            };
+
+            for(int i = 0; i < NumInvariantDim; i++)
+                if(invariant_lengths_[i] != bnScaleBiasMeanVarLengths_[i])
+                    throw std::runtime_error("Invalid lengths parameters!");
+
+            for(int j = 0, i = 0; j < NumInvariantDim; j++)
+            {
+                int dim                 = invariantDims_[j];
+                x_invariant_strides_[i] = xStrides[dim];
+                y_invariant_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims_[j];
+                x_reduce_strides_[i] = xStrides[dim];
+                y_reduce_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+            reduce_index_set_    = get_index_set<NumBatchNormReduceDim>(reduce_lengths_);
+
+            epsilon_       = type_convert<AccDataType>(epsilon);
+            averageFactor_ = type_convert<AccDataType>(averageFactor);
+
+            resultSave    = (resultSaveMean != nullptr && resultSaveInvVariance != nullptr);
+            resultRunning = (resultRunningMean != nullptr && resultRunningVariance != nullptr);
+        }
+
+        std::array<int, NumBatchNormReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumBatchNormReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths_;
+        const std::array<index_t, NumInvariantDim> bnScaleStrides_;
+        const std::array<index_t, NumInvariantDim> bnBiasStrides_;
+        const std::array<index_t, NumInvariantDim> bnMeanVarStrides_;
+
+        std::array<index_t, NumInvariantDim> x_invariant_strides_;
+        std::array<index_t, NumInvariantDim> y_invariant_strides_;
+        std::array<index_t, NumBatchNormReduceDim> x_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> y_reduce_strides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* bnScale_;
+        const BiasDataType* bnBias_;
+        const YElementwiseOp y_elementwise_op_;
+        YDataType* p_y_;
+
+        MeanVarDataType* resultSaveMean_;
+        MeanVarDataType* resultSaveInvVariance_;
+        MeanVarDataType* resultRunningMean_;
+        MeanVarDataType* resultRunningVariance_;
+
+        bool resultSave, resultRunning;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumBatchNormReduceDim>> reduce_index_set_;
+
+        AccDataType averageFactor_;
+        AccDataType epsilon_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            using ck::host_common::get_offset_from_index;
+
+            auto thread_reduce_func = [&](auto invariant_index) {
+                size_t x_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.x_invariant_strides_, invariant_index);
+                size_t y_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.y_invariant_strides_, invariant_index);
+                AccDataType mean     = type_convert<AccDataType>(0.0f);
+                AccDataType variance = type_convert<AccDataType>(0.0f);
+                int32_t curr_count   = 0;
+
+                // compute mean, variance using welford method
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+
+                    curr_count++;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType delta = x - mean;
+
+                    mean += delta / curr_count;
+
+                    AccDataType delta2 = x - mean;
+
+                    variance += delta * delta2;
+                };
+
+                // actual variance
+                variance = variance / curr_count;
+
+                // inv-variance defined as 1/sqrt(epsilon+variance)
+                AccDataType invVariance =
+                    type_convert<AccDataType>(1.0f) / ck::math::sqrt(arg.epsilon_ + variance);
+
+                // save the mean/inv-variance if required
+                if(arg.resultSave)
+                {
+                    size_t offset = get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_,
+                                                                           invariant_index);
+
+                    arg.resultSaveMean_[offset]        = type_convert<MeanVarDataType>(mean);
+                    arg.resultSaveInvVariance_[offset] = type_convert<MeanVarDataType>(invVariance);
+                };
+
+                // update the moving average if required
+                if(arg.resultRunning)
+                {
+                    size_t offset = get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_,
+                                                                           invariant_index);
+
+                    AccDataType oneMinusAverageFactor =
+                        type_convert<AccDataType>(1.0) - arg.averageFactor_;
+                    arg.resultRunningMean_[offset] = type_convert<MeanVarDataType>(
+                        type_convert<AccDataType>(arg.resultRunningMean_[offset]) *
+                            oneMinusAverageFactor +
+                        mean * arg.averageFactor_);
+                    arg.resultRunningVariance_[offset] = type_convert<MeanVarDataType>(
+                        arg.resultRunningVariance_[offset] * oneMinusAverageFactor +
+                        variance * arg.averageFactor_);
+                };
+
+                size_t scale_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnScaleStrides_, invariant_index);
+                size_t bias_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnBiasStrides_, invariant_index);
+
+                AccDataType scale = type_convert<AccDataType>(arg.bnScale_[scale_offset]);
+                AccDataType bias  = type_convert<AccDataType>(arg.bnBias_[bias_offset]);
+
+                // Normalization
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t y_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.y_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+                    auto y_offset = y_invariant_offset + y_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVariance;
+
+                    AccDataType y = scale * norm_x + bias;
+
+                    arg.y_elementwise_op_(y, y);
+
+                    arg.p_y_[y_offset] = type_convert<YDataType>(y);
+                };
+            };
+
+            std::size_t num_thread = std::thread::hardware_concurrency();
+            std::size_t work_per_thread =
+                (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t i_begin = it * work_per_thread;
+                std::size_t i_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread),
+                                             arg.invariant_index_set_.size());
+
+                auto f = [=] {
+                    for(std::size_t i = i_begin; i < i_end; ++i)
+                    {
+                        thread_reduce_func(arg.invariant_index_set_[i]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        (void)p_arg;
+
+        return (true);
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, 4> xyLengths,
+                        const std::array<index_t, 4> xStrides,
+                        const std::array<index_t, 4> yStrides,
+                        const std::array<int, 3> reduceDims,
+                        const std::array<index_t, 1> bnScaleBiasMeanVarLengths,
+                        const std::array<index_t, 1> bnScaleStrides,
+                        const std::array<index_t, 1> bnBiasStrides,
+                        const std::array<index_t, 1> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* bnScale,
+                        const void* bnBias,
+                        double epsilon,
+                        const YElementwiseOp y_elementwise_op,
+                        void* p_y,
+                        void* resultSaveMean,
+                        void* resultSaveInvVariance,
+                        double averageFactor,
+                        void* resultRunningMean,
+                        void* resultRunningVariance) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(bnScale),
+                                          static_cast<const BiasDataType*>(bnBias),
+                                          epsilon,
+                                          y_elementwise_op,
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<MeanVarDataType*>(resultSaveMean),
+                                          static_cast<MeanVarDataType*>(resultSaveInvVariance),
+                                          averageFactor,
+                                          static_cast<MeanVarDataType*>(resultRunningMean),
+                                          static_cast<MeanVarDataType*>(resultRunningVariance));
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_BatchNorm_Forward" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward_nhwc_c.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <iostream>
-#include <vector>
-#include <array>
-#include <algorithm>
-#include <thread>
-
-#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace host {
-
-template <typename InOutDataType, typename AccDataType>
-struct ReferenceBatchNormFwd_Input_N_H_W_C_Output_C : public device::DeviceBatchNormFwd<4, 3>
-{
-    struct Argument : public device::BaseArgument
-    {
-        Argument(const std::array<index_t, 4> xyLengths,
-                 const std::array<index_t, 4> xStrides,
-                 const std::array<index_t, 4> yStrides,
-                 const std::array<index_t, 1> bnScaleBiasMeanVarLengths,
-                 const std::array<index_t, 1> bnScaleBiasMeanVarStrides,
-                 const InOutDataType* p_x,
-                 const AccDataType* bnScale,
-                 const AccDataType* bnBias,
-                 InOutDataType* p_y,
-                 double exponentialAverageFactor,
-                 AccDataType* resultRunningMean,
-                 AccDataType* resultRunningVariance,
-                 double epsilon,
-                 AccDataType* resultSaveMean,
-                 AccDataType* resultSaveInvVariance)
-            : p_x_(p_x),
-              bnScale_(bnScale),
-              bnBias_(bnBias),
-              p_y_(p_y),
-              resultRunningMean_(resultRunningMean),
-              resultRunningVariance_(resultRunningVariance),
-              resultSaveMean_(resultSaveMean),
-              resultSaveInvVariance_(resultSaveInvVariance),
-              exponentialAverageFactor_(exponentialAverageFactor),
-              epsilon_(epsilon)
-        {
-            (void)xStrides;
-            (void)yStrides;
-            (void)bnScaleBiasMeanVarStrides;
-
-            if(xyLengths.size() != 4 || bnScaleBiasMeanVarLengths.size() != 1 ||
-               bnScaleBiasMeanVarLengths[0] != xyLengths[3])
-                throw std::runtime_error("Invalid tensor dimensions!");
-
-            n = xyLengths[0];
-            h = xyLengths[1];
-            w = xyLengths[2];
-            c = xyLengths[3];
-
-            resultSave    = (resultSaveMean != nullptr && resultSaveInvVariance != nullptr);
-            resultRunning = (resultRunningMean != nullptr && resultRunningVariance != nullptr);
-        }
-
-        const InOutDataType* p_x_;
-        const AccDataType* bnScale_;
-        const AccDataType* bnBias_;
-        InOutDataType* p_y_;
-
-        AccDataType* resultRunningMean_;
-        AccDataType* resultRunningVariance_;
-        AccDataType* resultSaveMean_;
-        AccDataType* resultSaveInvVariance_;
-
-        bool resultSave, resultRunning;
-
-        index_t n, h, w, c;
-
-        double exponentialAverageFactor_;
-        double epsilon_;
-    };
-
-    struct Invoker : public device::BaseInvoker
-    {
-        float Run(const Argument& arg)
-        {
-            auto thread_reduce_func = [&](auto iC) {
-                AccDataType reduceSize = type_convert<AccDataType>(arg.n) *
-                                         type_convert<AccDataType>(arg.h) *
-                                         type_convert<AccDataType>(arg.w);
-                index_t offset_C       = iC;
-                AccDataType mean       = type_convert<AccDataType>(0.0f);
-                AccDataType meansquare = type_convert<AccDataType>(0.0f);
-
-                // compute mean, meanquare, variance, invVariance
-                for(index_t iN = 0; iN < arg.n; iN++)
-                {
-                    index_t offset_N = iN * arg.h * arg.w * arg.c;
-                    for(index_t iH = 0; iH < arg.h; iH++)
-                    {
-                        index_t offset_H = iH * arg.w * arg.c;
-                        for(index_t iW = 0; iW < arg.w; iW++)
-                        {
-                            index_t offset_W = iW * arg.c;
-
-                            auto offset = offset_N + offset_H + offset_W + offset_C;
-
-                            AccDataType x = type_convert<AccDataType>(arg.p_x_[offset]);
-
-                            mean += x;
-                            meansquare += x * x;
-                        };
-                    }
-                };
-
-                mean       = mean / reduceSize;
-                meansquare = meansquare / reduceSize;
-
-                AccDataType variance = meansquare - mean * mean;
-                AccDataType invVariance =
-                    type_convert<AccDataType>(1.0f) /
-                    std::sqrt(type_convert<AccDataType>(arg.epsilon_) + variance);
-
-                // save the mean/invVariance if required
-                if(arg.resultSave)
-                {
-                    arg.resultSaveMean_[iC]        = mean;
-                    arg.resultSaveInvVariance_[iC] = invVariance;
-                };
-
-                // update the moving average if required
-                if(arg.resultRunning)
-                {
-                    arg.resultRunningMean_[iC] =
-                        arg.resultRunningMean_[iC] *
-                            type_convert<AccDataType>(1.0 - arg.exponentialAverageFactor_) +
-                        mean * arg.exponentialAverageFactor_;
-                    arg.resultRunningVariance_[iC] =
-                        arg.resultRunningVariance_[iC] *
-                            type_convert<AccDataType>(1.0 - arg.exponentialAverageFactor_) +
-                        variance * arg.exponentialAverageFactor_;
-                };
-
-                // Normalization
-                for(index_t iN = 0; iN < arg.n; iN++)
-                {
-                    index_t offset_N = iN * arg.h * arg.w * arg.c;
-                    for(index_t iH = 0; iH < arg.h; iH++)
-                    {
-                        index_t offset_H = iH * arg.w * arg.c;
-                        for(index_t iW = 0; iW < arg.w; iW++)
-                        {
-                            index_t offset_W = iW * arg.c;
-
-                            auto offset = offset_N + offset_H + offset_W + offset_C;
-
-                            AccDataType x = type_convert<AccDataType>(arg.p_x_[offset]);
-
-                            AccDataType norm_x =
-                                arg.bnScale_[iC] * (x - mean) * invVariance + arg.bnBias_[iC];
-
-                            arg.p_y_[offset] = type_convert<InOutDataType>(norm_x);
-                        };
-                    }
-                };
-            };
-
-            std::size_t num_thread      = std::thread::hardware_concurrency();
-            std::size_t work_per_thread = (arg.c + num_thread - 1) / num_thread;
-
-            std::vector<joinable_thread> threads(num_thread);
-
-            for(std::size_t it = 0; it < num_thread; ++it)
-            {
-                std::size_t ic_begin = it * work_per_thread;
-                std::size_t ic_end = std::min(static_cast<int>((it + 1) * work_per_thread), arg.c);
-
-                auto f = [=] {
-                    for(std::size_t ic = ic_begin; ic < ic_end; ++ic)
-                    {
-                        thread_reduce_func(ic);
-                    }
-                };
-
-                threads[it] = joinable_thread(f);
-            }
-
-            return (0.0f);
-        };
-
-        float Run(const device::BaseArgument* p_arg,
-                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
-        {
-            return Run(*dynamic_cast<const Argument*>(p_arg));
-        };
-    };
-
-    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
-    {
-        (void)p_arg;
-
-        return (true);
-    };
-
-    std::unique_ptr<device::BaseArgument>
-    MakeArgumentPointer(const std::array<index_t, 4> xyLengths,
-                        const std::array<index_t, 4> xStrides,
-                        const std::array<index_t, 4> yStrides,
-                        const std::array<index_t, 1> bnScaleBiasMeanVarLengths,
-                        const std::array<index_t, 1> bnScaleBiasMeanVarStrides,
-                        const void* p_x,
-                        const void* bnScale,
-                        const void* bnBias,
-                        void* p_y,
-                        double exponentialAverageFactor,
-                        void* resultRunningMean,
-                        void* resultRunningVariance,
-                        double epsilon,
-                        void* resultSaveMean,
-                        void* resultSaveInvVariance) override
-    {
-        return std::make_unique<Argument>(xyLengths,
-                                          xStrides,
-                                          yStrides,
-                                          bnScaleBiasMeanVarLengths,
-                                          bnScaleBiasMeanVarStrides,
-                                          static_cast<const InOutDataType*>(p_x),
-                                          static_cast<const AccDataType*>(bnScale),
-                                          static_cast<const AccDataType*>(bnBias),
-                                          static_cast<InOutDataType*>(p_y),
-                                          exponentialAverageFactor,
-                                          static_cast<AccDataType*>(resultRunningMean),
-                                          static_cast<AccDataType*>(resultRunningVariance),
-                                          epsilon,
-                                          static_cast<AccDataType*>(resultSaveMean),
-                                          static_cast<AccDataType*>(resultSaveInvVariance));
-    };
-
-    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
-    {
-        return std::make_unique<Invoker>();
-    };
-
-    std::string GetTypeString() const override
-    {
-        auto str = std::stringstream();
-
-        // clang-format off
-        str << "Reference_BatchNorm_Forward_NHWC_C<" << std::endl;
-        // clang-format on
-
-        return str.str();
-    }
-};
-
-} // namespace host
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct ReferenceBatchNormInfer : public device::DeviceBatchNormInfer<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     YElementwiseOp,
+                                                                     Rank,
+                                                                     NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                 const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* bnScale,
+                 const BiasDataType* bnBias,
+                 double epsilon,
+                 const YElementwiseOp y_elementwise_op,
+                 const MeanVarDataType* estimatedMean,
+                 const MeanVarDataType* estimatedVariance,
+                 YDataType* p_y)
+            : reduceDims_(reduceDims),
+              bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              bnScale_(bnScale),
+              bnBias_(bnBias),
+              y_elementwise_op_(y_elementwise_op),
+              estimatedMean_(estimatedMean),
+              estimatedVariance_(estimatedVariance),
+              p_y_(p_y)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            // get invariant_dims[] and invariant_lengths[]
+            for(int dim = 0, i = 0; dim < Rank; dim++)
+                if(std::none_of(
+                       reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                {
+                    invariantDims_[i]     = dim;
+                    invariant_lengths_[i] = xyLengths[dim];
+                    i++;
+                };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = xyLengths[dim];
+            };
+
+            // check invariant_lengths_ and bnScaleBiasMeanVarLengths
+            for(int i = 0; i < NumInvariantDim; i++)
+                if(invariant_lengths_[i] != bnScaleBiasMeanVarLengths_[i])
+                    throw std::runtime_error("Invalid lengths parameters!");
+
+            for(int j = 0, i = 0; j < NumInvariantDim; j++)
+            {
+                int dim                 = invariantDims_[j];
+                x_invariant_strides_[i] = xStrides[dim];
+                y_invariant_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims_[j];
+                x_reduce_strides_[i] = xStrides[dim];
+                y_reduce_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+            reduce_index_set_    = get_index_set<NumBatchNormReduceDim>(reduce_lengths_);
+
+            epsilon_ = type_convert<AccDataType>(epsilon);
+        }
+
+        std::array<int, NumBatchNormReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumBatchNormReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths_;
+        const std::array<index_t, NumInvariantDim> bnScaleStrides_;
+        const std::array<index_t, NumInvariantDim> bnBiasStrides_;
+        const std::array<index_t, NumInvariantDim> bnMeanVarStrides_;
+
+        std::array<index_t, NumInvariantDim> x_invariant_strides_;
+        std::array<index_t, NumInvariantDim> y_invariant_strides_;
+        std::array<index_t, NumBatchNormReduceDim> x_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> y_reduce_strides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* bnScale_;
+        const BiasDataType* bnBias_;
+        const YElementwiseOp y_elementwise_op_;
+
+        const MeanVarDataType* estimatedMean_;
+        const MeanVarDataType* estimatedVariance_;
+
+        YDataType* p_y_;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumBatchNormReduceDim>> reduce_index_set_;
+
+        AccDataType epsilon_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            using ck::host_common::get_offset_from_index;
+
+            auto thread_reduce_func = [&](auto invariant_index) {
+                size_t x_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.x_invariant_strides_, invariant_index);
+                size_t y_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.y_invariant_strides_, invariant_index);
+
+                size_t mean_variance_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_, invariant_index);
+
+                AccDataType mean     = arg.estimatedMean_[mean_variance_offset];
+                AccDataType variance = arg.estimatedVariance_[mean_variance_offset];
+
+                // inv-variance defined as 1/sqrt(epsilon+variance)
+                AccDataType invVariance =
+                    type_convert<AccDataType>(1.0f) / std::sqrt(arg.epsilon_ + variance);
+
+                size_t scale_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnScaleStrides_, invariant_index);
+                size_t bias_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnBiasStrides_, invariant_index);
+
+                AccDataType scale = type_convert<AccDataType>(arg.bnScale_[scale_offset]);
+                AccDataType bias  = type_convert<AccDataType>(arg.bnBias_[bias_offset]);
+
+                // normalization
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t y_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.y_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+                    auto y_offset = y_invariant_offset + y_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVariance;
+
+                    AccDataType y = scale * norm_x + bias;
+
+                    arg.y_elementwise_op_(y, y);
+
+                    arg.p_y_[y_offset] = type_convert<YDataType>(y);
+                };
+            };
+
+            std::size_t num_thread = std::thread::hardware_concurrency();
+            std::size_t work_per_thread =
+                (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t i_begin = it * work_per_thread;
+                std::size_t i_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread),
+                                             arg.invariant_index_set_.size());
+
+                auto f = [=] {
+                    for(std::size_t i = i_begin; i < i_end; ++i)
+                    {
+                        thread_reduce_func(arg.invariant_index_set_[i]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        (void)p_arg;
+
+        return (true);
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> yStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                        const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* bnScale,
+                        const void* bnBias,
+                        double epsilon,
+                        const YElementwiseOp y_elementwise_op,
+                        const void* estimatedMean,
+                        const void* estimatedVariance,
+                        void* p_y) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(bnScale),
+                                          static_cast<const BiasDataType*>(bnBias),
+                                          epsilon,
+                                          y_elementwise_op,
+                                          static_cast<const MeanVarDataType*>(estimatedMean),
+                                          static_cast<const MeanVarDataType*>(estimatedVariance),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_BatchNorm_Infer<" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer_nhwc_c.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <iostream>
-#include <vector>
-#include <array>
-#include <algorithm>
-
-#include "ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace host {
-
-template <typename InOutDataType, typename AccDataType>
-struct ReferenceBatchNormInfer_Input_N_H_W_C_Output_C : public device::DeviceBatchNormInfer<4, 3>
-{
-    struct Argument : public device::BaseArgument
-    {
-        Argument(const std::array<index_t, 4> xyLengths,
-                 const std::array<index_t, 4> xStrides,
-                 const std::array<index_t, 4> yStrides,
-                 const std::array<index_t, 1> bnScaleBiasMeanVarLengths,
-                 const std::array<index_t, 1> bnScaleBiasMeanVarStrides,
-                 const InOutDataType* p_x,
-                 const AccDataType* bnScale,
-                 const AccDataType* bnBias,
-                 double epsilon,
-                 const AccDataType* estimatedMean,
-                 const AccDataType* estimatedVariance,
-                 InOutDataType* p_y)
-            : p_x_(p_x),
-              bnScale_(bnScale),
-              bnBias_(bnBias),
-              epsilon_(epsilon),
-              estimatedMean_(estimatedMean),
-              estimatedVariance_(estimatedVariance),
-              p_y_(p_y)
-        {
-            (void)xStrides;
-            (void)yStrides;
-            (void)bnScaleBiasMeanVarStrides;
-
-            if(xyLengths.size() != 4 || bnScaleBiasMeanVarLengths.size() != 1 ||
-               bnScaleBiasMeanVarLengths[0] != xyLengths[3])
-                throw std::runtime_error("Invalid tensor dimensions!");
-
-            n = xyLengths[0];
-            h = xyLengths[1];
-            w = xyLengths[2];
-            c = xyLengths[3];
-        }
-
-        const InOutDataType* p_x_;
-        const AccDataType* bnScale_;
-        const AccDataType* bnBias_;
-
-        double epsilon_;
-
-        const AccDataType* estimatedMean_;
-        const AccDataType* estimatedVariance_;
-
-        InOutDataType* p_y_;
-
-        index_t n, h, w, c;
-    };
-
-    struct Invoker : public device::BaseInvoker
-    {
-        float Run(const Argument& arg)
-        {
-            auto thread_reduce_func = [&](auto iC) {
-                index_t offset_C     = iC;
-                AccDataType mean     = arg.estimatedMean_[offset_C];
-                AccDataType variance = arg.estimatedVariance_[offset_C];
-
-                AccDataType invVariance =
-                    type_convert<AccDataType>(1.0f) /
-                    std::sqrt(type_convert<AccDataType>(arg.epsilon_) + variance);
-
-                // Normalization
-                for(index_t iN = 0; iN < arg.n; iN++)
-                {
-                    index_t offset_N = iN * arg.h * arg.w * arg.c;
-                    for(index_t iH = 0; iH < arg.h; iH++)
-                    {
-                        index_t offset_H = iH * arg.w * arg.c;
-                        for(index_t iW = 0; iW < arg.w; iW++)
-                        {
-                            index_t offset_W = iW * arg.c;
-
-                            auto offset = offset_N + offset_H + offset_W + offset_C;
-
-                            AccDataType x = type_convert<AccDataType>(arg.p_x_[offset]);
-
-                            AccDataType norm_x =
-                                arg.bnScale_[iC] * (x - mean) * invVariance + arg.bnBias_[iC];
-
-                            arg.p_y_[offset] = type_convert<InOutDataType>(norm_x);
-                        };
-                    }
-                };
-            };
-
-            std::size_t num_thread      = std::thread::hardware_concurrency();
-            std::size_t work_per_thread = (arg.c + num_thread - 1) / num_thread;
-
-            std::vector<joinable_thread> threads(num_thread);
-
-            for(std::size_t it = 0; it < num_thread; ++it)
-            {
-                std::size_t ic_begin = it * work_per_thread;
-                std::size_t ic_end = std::min(static_cast<int>((it + 1) * work_per_thread), arg.c);
-
-                auto f = [=] {
-                    for(std::size_t ic = ic_begin; ic < ic_end; ++ic)
-                    {
-                        thread_reduce_func(ic);
-                    }
-                };
-
-                threads[it] = joinable_thread(f);
-            }
-
-            return (0.0f);
-        };
-
-        float Run(const device::BaseArgument* p_arg,
-                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
-        {
-            return Run(*dynamic_cast<const Argument*>(p_arg));
-        };
-    };
-
-    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
-    {
-        (void)p_arg;
-
-        return (true);
-    };
-
-    std::unique_ptr<device::BaseArgument>
-    MakeArgumentPointer(const std::array<index_t, 4> xyLengths,
-                        const std::array<index_t, 4> xStrides,
-                        const std::array<index_t, 4> yStrides,
-                        const std::array<index_t, 1> bnScaleBiasMeanVarLengths,
-                        const std::array<index_t, 1> bnScaleBiasMeanVarStrides,
-                        const void* p_x,
-                        const void* bnScale,
-                        const void* bnBias,
-                        double epsilon,
-                        const void* estimatedMean,
-                        const void* estimatedVariance,
-                        void* p_y) override
-    {
-        return std::make_unique<Argument>(xyLengths,
-                                          xStrides,
-                                          yStrides,
-                                          bnScaleBiasMeanVarLengths,
-                                          bnScaleBiasMeanVarStrides,
-                                          static_cast<const InOutDataType*>(p_x),
-                                          static_cast<const AccDataType*>(bnScale),
-                                          static_cast<const AccDataType*>(bnBias),
-                                          epsilon,
-                                          static_cast<const AccDataType*>(estimatedMean),
-                                          static_cast<const AccDataType*>(estimatedVariance),
-                                          static_cast<InOutDataType*>(p_y));
-    };
-
-    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
-    {
-        return std::make_unique<Invoker>();
-    };
-
-    std::string GetTypeString() const override
-    {
-        auto str = std::stringstream();
-
-        // clang-format off
-        str << "Reference_BatchNorm_Forward_NHWC_C<" << std::endl;
-        // clang-format on
-
-        return str.str();
-    }
-};
-
-} // namespace host
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp

@@ -131,17 +131,22 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
             else if constexpr(NDimSpatial == 2)
             {
                 auto f_kcyx = [&](auto g, auto k, auto c, auto y, auto x) {
+                    std::size_t N = arg.output_.GetLengths()[1];
+
+                    std::size_t Ho = arg.output_.GetLengths()[3];
+                    std::size_t Wo = arg.output_.GetLengths()[4];
+
                     float v_acc = 0;
 
-                    for(std::size_t n = 0; n < arg.output_.GetLengths()[1]; ++n)
+                    for(std::size_t n = 0; n < N; ++n)
                     {
-                        for(std::size_t ho = 0; ho < arg.output_.GetLengths()[3]; ++ho)
+                        for(std::size_t ho = 0; ho < Ho; ++ho)
                         {
                             auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
                                       static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
                                       static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
 
-                            for(std::size_t wo = 0; wo < arg.output_.GetLengths()[4]; ++wo)
+                            for(std::size_t wo = 0; wo < Wo; ++wo)
                             {
                                 auto wi =
                                     static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +

library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp

@@ -44,8 +44,8 @@ struct ReferenceGemmLayernorm : public device::BaseOperator
         size_t M = acc.mDesc.GetLengths()[0];
         size_t N = acc.mDesc.GetLengths()[1];
 
-        Tensor<ComputeDataType> avg_acc_sq(HostTensorDescriptor(std::vector<size_t>({M})));
-        Tensor<ComputeDataType> avg_acc(HostTensorDescriptor(std::vector<size_t>({M})));
+        Tensor<ComputeDataType> avg_acc_sq({M});
+        Tensor<ComputeDataType> avg_acc({M});
         Tensor<ComputeDataType> acc_layernorm(acc);
 
         // reduce N dim

library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp

@@ -90,11 +90,15 @@ struct ReferenceLayernorm : public device::BaseOperator
 
             for(int m = 0; m < M; ++m)
             {
+                AccDataType divisor =
+                    static_cast<AccDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_);
+
                 for(int n = 0; n < N; ++n)
                 {
-                    auto x_val       = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
-                    auto y_val       = (x_val - mean(m)) / sqrt(var(m) + arg.epsilon_);
-                    y_val            = (y_val * arg.gamma_n_(n)) + arg.beta_n_(n);
+                    auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
+                    auto y_val = (x_val - mean(m)) * divisor;
+                    y_val      = (y_val * arg.gamma_n_(n)) + arg.beta_n_(n);
+                    arg.acc_elementwise_op_(y_val, y_val);
                     arg.y_m_n_(m, n) = ck::type_convert<YDataType>(y_val);
                 }
             }

library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <thread>
+
+#include "ck/ck.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex>
+struct ReferenceReduce : public device::DeviceReduce<InDataType,
+                                                     AccDataType,
+                                                     OutDataType,
+                                                     Rank,
+                                                     NumReduceDim,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     PropagateNan,
+                                                     OutputIndex>
+{
+    using IndexDataType = int32_t;
+
+    static constexpr int NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> inLengths,
+                 const std::array<index_t, Rank> inStrides,
+                 const std::array<index_t, NumDstDim> outLengths,
+                 const std::array<index_t, NumDstDim> outStrides,
+                 const std::array<int, NumReduceDim> reduceDims,
+                 double alpha,
+                 double beta,
+                 const InDataType* in_host,
+                 OutDataType* out_host,
+                 IndexDataType* out_index_host,
+                 const InElementwiseOperation in_elementwise_op,
+                 const AccElementwiseOperation acc_elementwise_op)
+            : reduceDims_(reduceDims),
+              outLengths_(outLengths),
+              outStrides_(outStrides),
+              in_host_(in_host),
+              out_host_(out_host),
+              out_index_host_(out_index_host),
+              in_elementwise_op_(in_elementwise_op),
+              acc_elementwise_op_(acc_elementwise_op)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                // get invariant_dims[] and invariant_lengths[]
+                for(int dim = 0, i = 0; dim < Rank; dim++)
+                    if(std::none_of(
+                           reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                    {
+                        invariantDims_[i]     = dim;
+                        invariant_lengths_[i] = inLengths[dim];
+                        i++;
+                    };
+            };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = inLengths[dim];
+            };
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                // check invariant_lengths_ and outLengths
+                for(int i = 0; i < NumInvariantDim; i++)
+                    if(invariant_lengths_[i] != outLengths_[i])
+                        throw std::runtime_error("Invalid lengths parameters!");
+            }
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                for(int j = 0, i = 0; j < NumInvariantDim; j++)
+                {
+                    int dim                  = invariantDims_[j];
+                    in_invariant_strides_[i] = inStrides[dim];
+                    i++;
+                };
+            };
+
+            for(int j = 0, i = 0; j < NumReduceDim; j++)
+            {
+                int dim               = reduceDims_[j];
+                in_reduce_strides_[i] = inStrides[dim];
+                i++;
+            };
+
+            if constexpr(NumInvariantDim > 0)
+                invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+
+            reduce_index_set_ = get_index_set<NumReduceDim>(reduce_lengths_);
+
+            alpha_ = type_convert<AccDataType>(alpha);
+            beta_  = type_convert<AccDataType>(beta);
+        };
+
+        const std::array<int, NumReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumDstDim> outLengths_;
+        const std::array<index_t, NumDstDim> outStrides_;
+
+        std::array<index_t, NumInvariantDim> in_invariant_strides_;
+        std::array<index_t, NumReduceDim> in_reduce_strides_;
+
+        const InDataType* in_host_;
+        OutDataType* out_host_;
+        IndexDataType* out_index_host_;
+        const InElementwiseOperation in_elementwise_op_;
+        const AccElementwiseOperation acc_elementwise_op_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumReduceDim>> reduce_index_set_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            ignore = stream_config;
+
+            using ck::float_equal_one;
+            using ck::float_equal_zero;
+            using ck::type_convert;
+            using ck::host_common::get_index_set;
+            using ck::host_common::get_offset_from_index;
+
+            if constexpr(OutputIndex)
+            {
+                using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                                ReduceOperation,
+                                                                                AccDataType,
+                                                                                IndexDataType>;
+
+                if constexpr(NumInvariantDim == 0)
+                {
+                    AccDataType accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+                    IndexDataType accuIndex = 0;
+
+                    for(std::size_t i = 0; i < arg.reduce_index_set_.size(); i++)
+                    {
+                        auto in_offset = get_offset_from_index<NumReduceDim>(
+                            arg.in_reduce_strides_, arg.reduce_index_set_[i]);
+
+                        auto currVal = type_convert<AccDataType>(arg.in_host_[in_offset]);
+
+                        arg.in_elementwise_op_(currVal, currVal);
+
+                        auto currIndex = static_cast<IndexDataType>(i);
+
+                        Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                    };
+
+                    arg.acc_elementwise_op_(accuVal, accuVal);
+
+                    if(!float_equal_one{}(arg.alpha_))
+                        accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                    if(!float_equal_zero{}(arg.beta_))
+                        accuVal += type_convert<AccDataType>(arg.out_host_[0]) *
+                                   type_convert<AccDataType>(arg.beta_);
+
+                    arg.out_host_[0]       = type_convert<OutDataType>(accuVal);
+                    arg.out_index_host_[0] = accuIndex;
+                }
+                else
+                {
+                    auto thread_reduce_func = [&](auto invariant_index) {
+                        AccDataType accuVal =
+                            ReduceOperation::template GetIdentityValue<AccDataType>();
+                        IndexDataType accuIndex = 0;
+
+                        auto in_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                            arg.in_invariant_strides_, invariant_index);
+
+                        for(std::size_t i = 0; i < arg.reduce_index_set_.size(); i++)
+                        {
+                            auto in_reduce_offset = get_offset_from_index<NumReduceDim>(
+                                arg.in_reduce_strides_, arg.reduce_index_set_[i]);
+
+                            auto currVal = type_convert<AccDataType>(
+                                arg.in_host_[in_invariant_offset + in_reduce_offset]);
+
+                            arg.in_elementwise_op_(currVal, currVal);
+
+                            auto currIndex = static_cast<IndexDataType>(i);
+
+                            Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                        };
+
+                        arg.acc_elementwise_op_(accuVal, accuVal);
+
+                        if(!float_equal_one{}(arg.alpha_))
+                            accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                        auto dst_offset = get_offset_from_index<NumInvariantDim>(arg.outStrides_,
+                                                                                 invariant_index);
+
+                        if(!float_equal_zero{}(arg.beta_))
+                            accuVal += type_convert<AccDataType>(arg.out_host_[dst_offset]) *
+                                       type_convert<AccDataType>(arg.beta_);
+
+                        arg.out_host_[dst_offset]       = type_convert<OutDataType>(accuVal);
+                        arg.out_index_host_[dst_offset] = accuIndex;
+                    };
+
+                    std::size_t num_thread = std::thread::hardware_concurrency();
+
+                    std::size_t work_per_thread =
+                        (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+                    std::vector<joinable_thread> threads(num_thread);
+
+                    for(std::size_t it = 0; it < num_thread; ++it)
+                    {
+                        std::size_t i_begin = it * work_per_thread;
+                        std::size_t i_end =
+                            std::min((it + 1) * work_per_thread, arg.invariant_index_set_.size());
+
+                        auto f = [=] {
+                            for(std::size_t i = i_begin; i < i_end; i++)
+                            {
+                                thread_reduce_func(arg.invariant_index_set_[i]);
+                            }
+                        };
+
+                        threads[it] = joinable_thread(f);
+                    }
+                };
+            }
+            else
+            {
+                using Accumulation =
+                    ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+                if constexpr(NumInvariantDim == 0)
+                {
+                    AccDataType accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+                    for(const auto& reduce_index : arg.reduce_index_set_)
+                    {
+                        auto in_offset = get_offset_from_index<NumReduceDim>(arg.in_reduce_strides_,
+                                                                             reduce_index);
+
+                        auto currVal = type_convert<AccDataType>(arg.in_host_[in_offset]);
+
+                        arg.in_elementwise_op_(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal);
+                    };
+
+                    arg.acc_elementwise_op_(accuVal, accuVal);
+
+                    if(!float_equal_one{}(arg.alpha_))
+                        accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                    if(!float_equal_zero{}(arg.beta_))
+                        accuVal += type_convert<AccDataType>(arg.out_host_[0]) *
+                                   type_convert<AccDataType>(arg.beta_);
+
+                    arg.out_host_[0] = type_convert<OutDataType>(accuVal);
+                }
+                else
+                {
+                    auto thread_reduce_func = [&](auto invariant_index) {
+                        AccDataType accuVal =
+                            ReduceOperation::template GetIdentityValue<AccDataType>();
+
+                        auto in_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                            arg.in_invariant_strides_, invariant_index);
+
+                        for(const auto& reduce_index : arg.reduce_index_set_)
+                        {
+                            auto in_reduce_offset = get_offset_from_index<NumReduceDim>(
+                                arg.in_reduce_strides_, reduce_index);
+
+                            auto currVal = type_convert<AccDataType>(
+                                arg.in_host_[in_invariant_offset + in_reduce_offset]);
+
+                            arg.in_elementwise_op_(currVal, currVal);
+
+                            Accumulation::Calculate(accuVal, currVal);
+                        };
+
+                        arg.acc_elementwise_op_(accuVal, accuVal);
+
+                        if(!float_equal_one{}(arg.alpha_))
+                            accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                        auto dst_offset = get_offset_from_index<NumInvariantDim>(arg.outStrides_,
+                                                                                 invariant_index);
+
+                        if(!float_equal_zero{}(arg.beta_))
+                            accuVal += type_convert<AccDataType>(arg.out_host_[dst_offset]) *
+                                       type_convert<AccDataType>(arg.beta_);
+
+                        arg.out_host_[dst_offset] = type_convert<OutDataType>(accuVal);
+                    };
+
+                    std::size_t num_thread = std::thread::hardware_concurrency();
+
+                    std::size_t work_per_thread =
+                        (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+                    std::vector<joinable_thread> threads(num_thread);
+
+                    for(std::size_t it = 0; it < num_thread; ++it)
+                    {
+                        std::size_t i_begin = it * work_per_thread;
+                        std::size_t i_end =
+                            std::min((it + 1) * work_per_thread, arg.invariant_index_set_.size());
+
+                        auto f = [=] {
+                            for(std::size_t i = i_begin; i < i_end; i++)
+                            {
+                                thread_reduce_func(arg.invariant_index_set_[i]);
+                            }
+                        };
+
+                        threads[it] = joinable_thread(f);
+                    }
+                };
+            };
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        ignore = p_arg;
+
+        return true;
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumDstDim> outLengths,
+                        const std::array<index_t, NumDstDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        double alpha,
+                        double beta,
+                        const void* in_host,
+                        const void* in_index_host,
+                        void* out_host,
+                        void* out_index_host,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) override
+    {
+        ignore = in_index_host;
+
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStrides,
+                                          reduceDims,
+                                          alpha,
+                                          beta,
+                                          static_cast<const InDataType*>(in_host),
+                                          static_cast<OutDataType*>(out_host),
+                                          static_cast<IndexDataType*>(out_index_host),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_Reduce<" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp

@@ -24,11 +24,14 @@ struct ReferenceSoftmax : public device::BaseOperator
     {
         Argument(const Tensor<InDataType>& in,
                  Tensor<OutDataType>& out,
-                 AccDataType alpha,
-                 AccDataType beta,
+                 double alpha,
+                 double beta,
                  const std::vector<index_t> sm_reduce_dims)
-            : in_(in), out_(out), alpha_(alpha), beta_(beta), sm_reduce_dims_(sm_reduce_dims)
+            : in_(in), out_(out), sm_reduce_dims_(sm_reduce_dims)
         {
+            alpha_ = static_cast<AccDataType>(alpha);
+            beta_  = static_cast<AccDataType>(beta);
+
             // std::cout << "debug: scalar dims: ";
             for(size_t i = 0; i < in.mDesc.GetNumOfDimension(); i++)
             {
@@ -60,6 +63,12 @@ struct ReferenceSoftmax : public device::BaseOperator
             {
                 scalar_lengths.push_back(arg.in_.mDesc.GetLengths()[dim]);
             }
+            // max and sum reduction with final reduced values of dim=0 is a scalar so give it
+            // appropriate lengths of {1}
+            if(arg.sm_scalar_dims_.size() == 0)
+            {
+                scalar_lengths.push_back(1);
+            }
 
             Tensor<AccDataType> reduce_max(scalar_lengths);
             reduce_max.GenerateTensorValue(
@@ -67,6 +76,9 @@ struct ReferenceSoftmax : public device::BaseOperator
             Tensor<AccDataType> reduce_sum(scalar_lengths);
             reduce_sum.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0});
 
+            // when final reduced values is of dim=0, the index will be transformed into empty
+            // std::vector which is actually a valid input for Tensor::operator(std::vector) and
+            // internally accesses 0'th element
             auto to_sm_scalar_idx = [&](auto idx) {
                 std::vector<size_t> sm_scalar_idx;
                 for(index_t dim : arg.sm_scalar_dims_)
@@ -77,8 +89,8 @@ struct ReferenceSoftmax : public device::BaseOperator
             };
 
             arg.in_.ForEach([&](auto& self, auto idx) {
-                reduce_max(to_sm_scalar_idx(idx)) = std::max(reduce_max(to_sm_scalar_idx(idx)),
-                                                             static_cast<AccDataType>(self(idx)));
+                reduce_max(to_sm_scalar_idx(idx)) = std::max(
+                    reduce_max(to_sm_scalar_idx(idx)), ck::type_convert<AccDataType>(self(idx)));
             });
 
             // LogRangeAsType<float>(std::cout << "reduce_max: ", reduce_max.mData, ",") <<
@@ -87,7 +99,7 @@ struct ReferenceSoftmax : public device::BaseOperator
             Tensor<AccDataType> in_stable(arg.in_.mDesc);
             in_stable.ForEach([&](auto& self, auto idx) {
                 // numerator = exp(x - max(x))
-                self(idx) = std::exp(static_cast<AccDataType>(arg.in_(idx)) -
+                self(idx) = std::exp(ck::type_convert<AccDataType>(arg.in_(idx)) -
                                      reduce_max(to_sm_scalar_idx(idx)));
             });
 
@@ -102,8 +114,10 @@ struct ReferenceSoftmax : public device::BaseOperator
             // std::endl;
 
             arg.out_.ForEach([&](auto& self, auto idx) {
-                self(idx) = arg.alpha_ * in_stable(idx) / reduce_sum(to_sm_scalar_idx(idx)) +
-                            arg.beta_ * self(idx);
+                AccDataType temp_result =
+                    arg.alpha_ * in_stable(idx) / reduce_sum(to_sm_scalar_idx(idx)) +
+                    arg.beta_ * self(idx);
+                self(idx) = ck::type_convert<OutDataType>(temp_result);
             });
 
             // LogRangeAsType<float>(std::cout << "out: ", arg.out_.mData, ",") << std::endl;
@@ -132,8 +146,8 @@ struct ReferenceSoftmax : public device::BaseOperator
 
     static auto MakeArgument(const Tensor<InDataType>& in,
                              Tensor<OutDataType>& out,
-                             AccDataType alpha,
-                             AccDataType beta,
+                             double alpha,
+                             double beta,
                              const std::vector<index_t> sm_reduce_dims)
     {
         return Argument{in, out, alpha, beta, sm_reduce_dims};

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -3,10 +3,10 @@
 
 #pragma once
 
-#include "ck/utility/data_type.hpp"
-#include "ck/utility/tuple.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -26,7 +26,9 @@ using Empty_Tuple = ck::Tuple<>;
 using F16_Tuple     = ck::Tuple<F16>;
 using F16_F16_Tuple = ck::Tuple<F16, F16>;
 
-using F32_Tuple = ck::Tuple<F32>;
+using F32_Tuple     = ck::Tuple<F32>;
+using I32_Tuple     = ck::Tuple<I32>;
+using I32_F32_Tuple = ck::Tuple<I32, F32>;
 
 // GEMM layout
 using Row = ck::tensor_layout::gemm::RowMajor;
@@ -75,13 +77,38 @@ using NWGK   = ck::tensor_layout::convolution::NWGK;
 using NHWGK  = ck::tensor_layout::convolution::NHWGK;
 using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
 
+//
+using GK          = ck::tensor_layout::convolution::G_K;
+using GK_Tuple    = ck::Tuple<GK>;
+using GK_GK_Tuple = ck::Tuple<GK, GK>;
+
 // pointwise functor
 using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using Relu           = ck::tensor_operation::element_wise::Relu;
 using Scale          = ck::tensor_operation::element_wise::Scale;
 using Bilinear       = ck::tensor_operation::element_wise::Bilinear;
 using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
+using AddReluAdd     = ck::tensor_operation::element_wise::AddReluAdd;
+using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
+using AddMultiply    = ck::tensor_operation::element_wise::AddMultiply;
+using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;
+
+template <typename Activation>
+using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;
+
+template <typename Activation>
+using Add_Activation_Mul_Clamp =
+    ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<Activation>;
+
+template <typename Activation>
+using Activation_Mul2_Clamp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<Activation>;
+
+template <typename Activation>
+using Add_Activation_Mul2_Clamp =
+    ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<Activation>;
 
-template <typename DeviceOp>
+template <typename DeviceOp, typename Tag = void>
 struct DeviceOperationInstanceFactory;
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance(
+    std::vector<std::unique_ptr<
+        DeviceBatchedContractionMultipleD<1,
+                                          2,
+                                          3,
+                                          1,
+                                          F16,
+                                          F16,
+                                          F16_Tuple,
+                                          F16,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::Add>>>& instances);
+
+// Contraction + add
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedContractionMultipleD<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Add>>
+{
+    using DeviceOp =
+        DeviceBatchedContractionMultipleD<NumDimG,
+                                          NumDimM,
+                                          NumDimN,
+                                          NumDimK,
+                                          ADataType,
+                                          BDataType,
+                                          ck::Tuple<DDataType>,
+                                          EDataType,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::Add>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::half_t> &&
+                     is_same_v<DDataType, ck::half_t> && is_same_v<EDataType, ck::half_t>)
+        {
+            if constexpr(NumDimG == 1 && NumDimM == 2 && NumDimN == 3 && NumDimK == 1)
+            {
+                add_device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<F16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<F16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<BF16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<BF16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          MaskingSpecialization MaskingSpec>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      Acc0BiasDataType,
+                                                                      ck::Tuple<>,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      ScaleAdd,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      MaskingSpec>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         ADataType,
+                                                         B0DataType,
+                                                         B1DataType,
+                                                         CDataType,
+                                                         Acc0BiasDataType,
+                                                         ck::Tuple<>,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         ScaleAdd,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         MaskingSpec>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t> &&
+                     Acc0BiasDataType::Size() == 1 &&
+                     is_same_v<tuple_element_t<0, Acc0BiasDataType>, half_t>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
+                          is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> &&
+                          Acc0BiasDataType::Size() == 1 &&
+                          is_same_v<tuple_element_t<0, Acc0BiasDataType>, BF16>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp

@@ -28,9 +28,26 @@ void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_g
                                                              F16,
                                                              PassThrough,
                                                              PassThrough,
+                                                             Scale,
                                                              PassThrough,
                                                              PassThrough,
-                                                             PassThrough>>>& instances);
+                                                             false>>>& instances);
+
+void add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             true>>>& instances);
 
 template <typename ALayout,
           typename B0Layout,
@@ -39,7 +56,8 @@ template <typename ALayout,
           typename ADataType,
           typename B0DataType,
           typename B1DataType,
-          typename CDataType>
+          typename CDataType,
+          bool MaskOutUpperTriangle>
 struct DeviceOperationInstanceFactory<
     ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm<ALayout,
                                                                B0Layout,
@@ -51,9 +69,10 @@ struct DeviceOperationInstanceFactory<
                                                                CDataType,
                                                                PassThrough,
                                                                PassThrough,
+                                                               Scale,
                                                                PassThrough,
                                                                PassThrough,
-                                                               PassThrough>>
+                                                               MaskOutUpperTriangle>>
 {
     using DeviceOp = DeviceBatchedGemmSoftmaxGemm<ALayout,
                                                   B0Layout,
@@ -65,9 +84,10 @@ struct DeviceOperationInstanceFactory<
                                                   CDataType,
                                                   PassThrough,
                                                   PassThrough,
+                                                  Scale,
                                                   PassThrough,
                                                   PassThrough,
-                                                  PassThrough>;
+                                                  MaskOutUpperTriangle>;
 
     static auto GetInstances()
     {
@@ -79,8 +99,16 @@ struct DeviceOperationInstanceFactory<
             if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
                          is_same_v<B1Layout, Row> && is_same_v<CLayout, Row>)
             {
-                add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
-                    op_ptrs);
+                if constexpr(MaskOutUpperTriangle)
+                {
+                    add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                        op_ptrs);
+                }
+                else
+                {
+                    add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                        op_ptrs);
+                }
             }
         }
         return op_ptrs;

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          MaskingSpecialization MaskingSpec>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      ck::Tuple<>,
+                                                                      ck::Tuple<>,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      Scale,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      MaskingSpec>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         ADataType,
+                                                         B0DataType,
+                                                         B1DataType,
+                                                         CDataType,
+                                                         ck::Tuple<>,
+                                                         ck::Tuple<>,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Scale,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         MaskingSpec>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
+                          is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_backward_rank_4_3_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F16, F32, F32, F32, F16, F32, F32, PassThrough, 4, 3>>>&);
+
+// FP32
+void add_device_batchnorm_backward_rank_4_3_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F32, F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
+
+// BF16
+void add_device_batchnorm_backward_rank_4_3_bf16_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<BF16, F32, F32, F32, BF16, F32, F32, PassThrough, 4, 3>>>&);
+
+// FP64
+void add_device_batchnorm_backward_rank_4_3_f64_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F64, F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&);
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                     DxDataType,
+                                                     DyDataType,
+                                                     AccDataType,
+                                                     ScaleDataType,
+                                                     DscaleDbiasDataType,
+                                                     MeanVarDataType,
+                                                     DyElementwiseOp,
+                                                     Rank,
+                                                     NumReduceDim>>
+{
+    using DeviceOp = DeviceBatchNormBwd<XDataType,
+                                        DxDataType,
+                                        DyDataType,
+                                        AccDataType,
+                                        ScaleDataType,
+                                        DscaleDbiasDataType,
+                                        MeanVarDataType,
+                                        DyElementwiseOp,
+                                        Rank,
+                                        NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<DxDataType, F32> &&
+                     is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                     is_same_v<ScaleDataType, F16> && is_same_v<DscaleDbiasDataType, F32> &&
+                     is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<DxDataType, F32> &&
+                          is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                          is_same_v<ScaleDataType, F32> && is_same_v<DscaleDbiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<DxDataType, F32> &&
+                          is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                          is_same_v<ScaleDataType, BF16> && is_same_v<DscaleDbiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<DxDataType, F64> &&
+                          is_same_v<DyDataType, F64> && is_same_v<AccDataType, F64> &&
+                          is_same_v<ScaleDataType, F64> && is_same_v<DscaleDbiasDataType, F64> &&
+                          is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_forward_rank_4_3_f16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F16, F16, F32, F16, F16, F32, PassThrough, 4, 3>>>&);
+
+// FP32
+void add_device_batchnorm_forward_rank_4_3_f32_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
+
+// BF16
+void add_device_batchnorm_forward_rank_4_3_bf16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<BF16, BF16, F32, BF16, BF16, F32, PassThrough, 4, 3>>>&);
+
+// FP64
+void add_device_batchnorm_forward_rank_4_3_f64_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&);
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                     YDataType,
+                                                     AccDataType,
+                                                     ScaleDataType,
+                                                     BiasDataType,
+                                                     MeanVarDataType,
+                                                     YElementwiseOp,
+                                                     Rank,
+                                                     NumReduceDim>>
+{
+    using DeviceOp = DeviceBatchNormFwd<XDataType,
+                                        YDataType,
+                                        AccDataType,
+                                        ScaleDataType,
+                                        BiasDataType,
+                                        MeanVarDataType,
+                                        YElementwiseOp,
+                                        Rank,
+                                        NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> &&
+                     is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F16> &&
+                     is_same_v<BiasDataType, F16> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> &&
+                          is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F32> &&
+                          is_same_v<BiasDataType, F32> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> &&
+                          is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, BF16> &&
+                          is_same_v<BiasDataType, BF16> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> &&
+                          is_same_v<AccDataType, F64> && is_same_v<ScaleDataType, F64> &&
+                          is_same_v<BiasDataType, F64> && is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_infer_rank_4_f16_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F16, F32, F32, F16, F16>,
+        ck::Tuple<F16>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// FP32
+void add_device_batchnorm_infer_rank_4_f32_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F32, F32, F32, F32, F32>,
+        ck::Tuple<F32>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// BF16
+void add_device_batchnorm_infer_rank_4_bf16_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<BF16, F32, F32, BF16, BF16>,
+        ck::Tuple<BF16>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// FP64
+void add_device_batchnorm_infer_rank_4_f64_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F64, F64, F64, F64, F64>,
+        ck::Tuple<F64>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+template <typename XDataType,
+          typename YDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          index_t Rank>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwise<
+    ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+    ck::Tuple<YDataType>,
+    ck::tensor_operation::element_wise::NormalizeInInfer,
+    Rank>>
+{
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+        ck::Tuple<YDataType>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        Rank>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> &&
+                     is_same_v<ScaleDataType, F16> && is_same_v<BiasDataType, F16> &&
+                     is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> &&
+                          is_same_v<ScaleDataType, F32> && is_same_v<BiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> &&
+                          is_same_v<ScaleDataType, BF16> && is_same_v<BiasDataType, BF16> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> &&
+                          is_same_v<ScaleDataType, F64> && is_same_v<BiasDataType, F64> &&
+                          is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp

@@ -101,6 +101,42 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough>>>& instances);
 
+// conv2d dl
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
 // conv3d backward data
 void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<3,
@@ -216,11 +252,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
                          is_same_v<OutDataType, float>)
             {
                 add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
             }
             else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
                               is_same_v<OutDataType, half_t>)
             {
                 add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
             }
             else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
                               is_same_v<WeiDataType, ck::bhalf_t> &&
@@ -232,6 +270,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
                               is_same_v<OutDataType, int8_t>)
             {
                 add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
             }
         }
         else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&

library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_weight.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <cstdlib>
-
-#include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/device_conv_bwd_weight.hpp"
-#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-// conv1d backward weight
-void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_bf16_f32_bf16_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<1,
-                                                    NWC,
-                                                    KXC,
-                                                    NWK,
-                                                    BF16,
-                                                    F32,
-                                                    BF16,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f16_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<1,
-                                                    NWC,
-                                                    KXC,
-                                                    NWK,
-                                                    F16,
-                                                    F16,
-                                                    F16,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-void add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f32_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<1,
-                                                    NWC,
-                                                    KXC,
-                                                    NWK,
-                                                    F32,
-                                                    F32,
-                                                    F32,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-// conv2d backward weight
-void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_bf16_f32_bf16_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<2,
-                                                    NHWC,
-                                                    KYXC,
-                                                    NHWK,
-                                                    BF16,
-                                                    F32,
-                                                    BF16,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<2,
-                                                    NHWC,
-                                                    KYXC,
-                                                    NHWK,
-                                                    F16,
-                                                    F16,
-                                                    F16,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-void add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<2,
-                                                    NHWC,
-                                                    KYXC,
-                                                    NHWK,
-                                                    F32,
-                                                    F32,
-                                                    F32,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-// conv3d backward weight
-void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_bf16_f32_bf16_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<3,
-                                                    NDHWC,
-                                                    KZYXC,
-                                                    NDHWK,
-                                                    BF16,
-                                                    F32,
-                                                    BF16,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<3,
-                                                    NDHWC,
-                                                    KZYXC,
-                                                    NDHWK,
-                                                    F16,
-                                                    F16,
-                                                    F16,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-void add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
-    std::vector<std::unique_ptr<DeviceConvBwdWeight<3,
-                                                    NDHWC,
-                                                    KZYXC,
-                                                    NDHWK,
-                                                    F32,
-                                                    F32,
-                                                    F32,
-                                                    PassThrough,
-                                                    PassThrough,
-                                                    PassThrough>>>& instances);
-
-template <ck::index_t NumDimSpatial,
-          typename InLayout,
-          typename WeiLayout,
-          typename OutLayout,
-          typename InDataType,
-          typename WeiDataType,
-          typename OutDataType>
-struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBwdWeight<
-    NumDimSpatial,
-    InLayout,
-    WeiLayout,
-    OutLayout,
-    InDataType,
-    WeiDataType,
-    OutDataType,
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough>>
-{
-    using DeviceOp = DeviceConvBwdWeight<NumDimSpatial,
-                                         InLayout,
-                                         WeiLayout,
-                                         OutLayout,
-                                         InDataType,
-                                         WeiDataType,
-                                         OutDataType,
-                                         ck::tensor_operation::element_wise::PassThrough,
-                                         ck::tensor_operation::element_wise::PassThrough,
-                                         ck::tensor_operation::element_wise::PassThrough>;
-
-    static auto GetInstances()
-    {
-        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-
-        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, NWC> && is_same_v<WeiLayout, KXC> &&
-                     is_same_v<OutLayout, NWK>)
-        {
-            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
-                         is_same_v<OutDataType, float>)
-            {
-                add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f32_instances(op_ptrs);
-            }
-            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
-                              is_same_v<OutDataType, half_t>)
-            {
-                add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_f16_instances(op_ptrs);
-            }
-            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
-                              is_same_v<OutDataType, ck::bhalf_t>)
-            {
-                add_device_conv1d_bwd_weight_xdl_nwc_kxc_nwk_bf16_f32_bf16_instances(op_ptrs);
-            }
-        }
-        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
-                          is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
-        {
-            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
-                         is_same_v<OutDataType, float>)
-            {
-                add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
-            }
-            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
-                              is_same_v<OutDataType, half_t>)
-            {
-                add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
-            }
-            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
-                              is_same_v<OutDataType, ck::bhalf_t>)
-            {
-                add_device_conv2d_bwd_weight_xdl_nhwc_kyxc_nhwk_bf16_f32_bf16_instances(op_ptrs);
-            }
-        }
-        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
-                          is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
-        {
-            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
-                         is_same_v<OutDataType, float>)
-            {
-                add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs);
-            }
-            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
-                              is_same_v<OutDataType, half_t>)
-            {
-                add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_f16_instances(op_ptrs);
-            }
-            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
-                              is_same_v<OutDataType, ck::bhalf_t>)
-            {
-                add_device_conv3d_bwd_weight_xdl_ndhwc_kzyxc_ndhwk_bf16_f32_bf16_instances(op_ptrs);
-            }
-        }
-
-        return op_ptrs;
-    }
-};
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp

@@ -7,7 +7,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
@@ -18,11 +18,8 @@ namespace device {
 namespace instance {
 
 using Normalize                            = ck::tensor_operation::element_wise::Normalize;
-using DeviceNormalizeFromMeanMeanSquarePtr = ck::tensor_operation::device::DeviceElementwiseBasePtr<
-    Tuple<half_t, float, float, half_t, half_t>,
-    Tuple<half_t>,
-    Normalize,
-    2>;
+using DeviceNormalizeFromMeanMeanSquarePtr = ck::tensor_operation::device::
+    DeviceElementwisePtr<Tuple<half_t, float, float, half_t, half_t>, Tuple<half_t>, Normalize, 2>;
 
 void add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(
     std::vector<DeviceNormalizeFromMeanMeanSquarePtr>& instances);