Merge branch 'develop' into gridwise_2d

include/ck/tensor_operation/gpu/grid/gridwise_layernorm_welford_variance.hpp → include/ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp

@@ -11,7 +11,7 @@
 
 namespace ck {
 
-// Y = LayerNorm(X, Beta, Gamma)
+// Y = Normalization(X, Beta, Gamma)
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
@@ -33,7 +33,7 @@ template <typename XDataType,
           index_t YDstVectorDim,
           index_t YDstVectorSize,
           bool SweepOnce>
-struct GridwiseLayernormWelfordVariance_mk_to_mk
+struct GridwiseNormalizationWelfordVariance_mk_to_mk
 {
     static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
                       (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),

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

@@ -60,6 +60,12 @@ 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 +73,9 @@ 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_)

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 {
@@ -28,6 +28,8 @@ using F16_F16_Tuple = ck::Tuple<F16, F16>;
 
 using F32_Tuple = ck::Tuple<F32>;
 
+using I32_Tuple = ck::Tuple<I32>;
+
 // GEMM layout
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -75,12 +77,24 @@ 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>;
+
 // 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;
 
+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 DeviceOp>
 struct DeviceOperationInstanceFactory;
 

library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.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_elementwise_normalization.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_elementwise_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<DeviceElementwiseNormalization<ck::Tuple<F16, F16>,
+                                                               F16,
+                                                               F16,
+                                                               F32,
+                                                               F16,
+                                                               element_wise::Add,
+                                                               PassThrough,
+                                                               2,
+                                                               1>>>&);
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwiseNormalization<
+    InDataTypeTuple,
+    GammaDataType,
+    BetaDataType,
+    F32,
+    YDataType,
+    ck::tensor_operation::element_wise::Add,
+    ck::tensor_operation::element_wise::PassThrough,
+    Rank,
+    NumReduceDim>>
+{
+    using DeviceOp = DeviceElementwiseNormalization<InDataTypeTuple,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    F32,
+                                                    YDataType,
+                                                    ck::tensor_operation::element_wise::Add,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    Rank,
+                                                    NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<GammaDataType, F16> && is_same_v<BetaDataType, F16> &&
+                     is_same_v<YDataType, F16>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_elementwise_normalization_rank_2_1_f16_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data.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 {
+
+// conv2d backward data
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdData<2,
+                                                         GNHWC,
+                                                         GKYXC,
+                                                         GNHWK,
+                                                         F16,
+                                                         F16,
+                                                         F16,
+                                                         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::DeviceGroupedConvBwdData<
+    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 = DeviceGroupedConvBwdData<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 == 2 && is_same_v<InLayout, GNHWC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                         is_same_v<OutDataType, F16>)
+            {
+                add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_bias_forward_perlayer_quantization.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_grouped_conv_fwd_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 {
+
+// grouped conv2d forward, GNHWC/GKYXC/GNHWK
+void add_device_conv2d_bias_perlayer_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      GNHWC,
+                                                      GKYXC,
+                                                      GK_TUPLE,
+                                                      GNHWK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_bias_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_TUPLE,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Activation_Mul_Clamp<Relu>>>>&
+        instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Add_Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   DsLayout,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   DsDataType,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Add_Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_TUPLE> &&
+                     is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<DsDataType, I32_Tuple> && is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                    add_device_conv2d_bias_perlayer_quantization_int8_instances(op_ptrs);
+                else if constexpr(is_same_v<Activation, Relu>)
+                    add_device_conv2d_bias_relu_perlayer_quantization_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_perlayer_quantization.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_grouped_conv_fwd_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 {
+
+// grouped conv2d forward, GNHWC/GKYXC/GNHWK
+void add_device_conv2d_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul_Clamp<Relu>>>>&
+        instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    Empty_Tuple,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    Empty_Tuple,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   Empty_Tuple,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   Empty_Tuple,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                    add_device_conv2d_perlayer_quantization_int8_instances(op_ptrs);
+                else if constexpr(is_same_v<Activation, Relu>)
+                    add_device_conv2d_relu_perlayer_quantization_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -8,20 +8,13 @@
 
 #include "ck/ck.hpp"
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
-#include "ck/utility/data_type.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
 
-using F16         = ck::half_t;
-using F32         = float;
-using PassThrough = ck::tensor_operation::element_wise::PassThrough;
-
 void add_device_softmax_f16_f16_rank3_instances(
     std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>&);
 void add_device_softmax_f16_f16_rank4_instances(
@@ -32,6 +25,11 @@ void add_device_softmax_f32_f32_rank3_instances(
 void add_device_softmax_f32_f32_rank4_instances(
     std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>&);
 
+void add_device_softmax_i8_i8_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>&);
+void add_device_softmax_i8_i8_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>&);
+
 template <typename InDataType, typename AccDataType, typename OutDataType, index_t Rank>
 struct DeviceOperationInstanceFactory<
     ck::tensor_operation::device::
@@ -60,6 +58,14 @@ struct DeviceOperationInstanceFactory<
             else if constexpr(Rank == 4)
                 add_device_softmax_f32_f32_rank4_instances(op_ptrs);
         }
+        else if constexpr(std::is_same_v<InDataType, I8> && std::is_same_v<AccDataType, F32> &&
+                          std::is_same_v<OutDataType, I8>)
+        {
+            if constexpr(Rank == 3)
+                add_device_softmax_i8_i8_rank3_instances(op_ptrs);
+            else if constexpr(Rank == 4)
+                add_device_softmax_i8_i8_rank4_instances(op_ptrs);
+        }
 
         return op_ptrs;
     }

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+void add_device_softmax_f16_f16_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f16_f16_instances = std::tuple<
+    // clang-format off
+    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+    // fallback kernel
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,              1>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,               8,              8>,
+    // Reduction on middle dimensions
+    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               1,              1>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               8,              4>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+void add_device_softmax_f32_f32_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck