Merge remote-tracking branch 'origin/develop' into bwroblew/warp_wise_dpp8

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

@@ -256,10 +256,12 @@ struct ReferencePoolingFwd : public device::BaseOperator
 
                     for(ck::index_t y = 0; y < arg.window_spatial_lengths_[0]; ++y)
                     {
-                        ck::index_t hi = ho * arg.window_strides_[0] + y - arg.in_left_pads_[0];
+                        ck::index_t hi = ho * arg.window_strides_[0] +
+                                         y * arg.window_dilations_[0] - arg.in_left_pads_[0];
                         for(ck::index_t x = 0; x < arg.window_spatial_lengths_[1]; ++x)
                         {
-                            ck::index_t wi = wo * arg.window_strides_[1] + x - arg.in_left_pads_[1];
+                            ck::index_t wi = wo * arg.window_strides_[1] +
+                                             x * arg.window_dilations_[1] - arg.in_left_pads_[1];
                             if(hi >= 0 &&
                                hi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[2]) &&
                                wi >= 0 &&

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

@@ -32,6 +32,8 @@ using F32_Tuple     = ck::Tuple<F32>;
 using I32_Tuple     = ck::Tuple<I32>;
 using I32_F32_Tuple = ck::Tuple<I32, F32>;
 
+using F32_F32_Tuple = ck::Tuple<F32, F32>;
+
 // GEMM layout
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -95,9 +97,11 @@ 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 MultiplyAdd    = ck::tensor_operation::element_wise::MultiplyAdd;
 using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;
 using Gelu           = ck::tensor_operation::element_wise::Gelu;
 using Swish          = ck::tensor_operation::element_wise::Swish;
+using Add            = ck::tensor_operation::element_wise::Add;
 
 template <typename Activation>
 using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;

library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#ifdef CK_ENABLE_FP16
+void add_device_avgpool_bwd_ndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F16, F16, NDHWC, NDHWC>>>&);
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_avgpool_bwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, BF16, BF16, NDHWC, NDHWC>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_avgpool_bwd_ndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F32, F32, NDHWC, NDHWC>>>&);
+#endif
+template <typename DOutDataType, typename DInDataType, typename InLayout, typename OutLayout>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::
+        DeviceAvgPoolBwd<3, DOutDataType, DInDataType, InLayout, OutLayout>>
+{
+    using DeviceOp = DeviceAvgPoolBwd<3, DOutDataType, DInDataType, InLayout, OutLayout>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(is_same_v<InLayout, NDHWC> && is_same_v<OutLayout, NDHWC>)
+        {
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<DOutDataType, F16> && is_same_v<DInDataType, F16>)
+                add_device_avgpool_bwd_ndhwc_f16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_BF16
+            else if constexpr(is_same_v<DOutDataType, BF16> && is_same_v<DInDataType, BF16>)
+                add_device_avgpool_bwd_ndhwc_bf16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_FP32
+            else if constexpr(is_same_v<DOutDataType, F32> && is_same_v<DInDataType, F32>)
+                add_device_avgpool_bwd_ndhwc_f32_instances(op_ptrs);
+#endif
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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_gemm_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_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+// GEMM + Multiply + Add
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::MultiplyAdd>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout, D1Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType, D1DataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::MultiplyAdd>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
+                     is_same_v<D0DataType, float> && is_same_v<D1DataType, float> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// fp16_output
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Row,
+                                                         Row_Tuple,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         F32_Tuple,
+                                                         F16,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances);
+
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Col,
+                                                         Row_Tuple,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         F32_Tuple,
+                                                         F16,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances);
+
+// fp32_output
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Row,
+                                                         Row_Tuple,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         F32_Tuple,
+                                                         F32,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances);
+
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Col,
+                                                         Row_Tuple,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         F32_Tuple,
+                                                         F32,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                           BLayout,
+                                                           Row_Tuple,
+                                                           ELayout,
+                                                           ADataType,
+                                                           BDataType,
+                                                           F32_Tuple,
+                                                           EDataType,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Add>>
+{
+    using DeviceOp = DeviceGroupedGemmFixedNK<ALayout,
+                                              BLayout,
+                                              Row_Tuple,
+                                              ELayout,
+                                              ADataType,
+                                              BDataType,
+                                              F32_Tuple,
+                                              EDataType,
+                                              PassThrough,
+                                              PassThrough,
+                                              Add>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        // fp16_output
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+        }
+
+        // fp32_output
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instances(op_ptrs);
+            }
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instances(op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#ifdef CK_ENABLE_FP16
+void add_device_maxpool_bwd_f16_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<F16, I32, F16>>>&);
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_maxpool_bwd_bf16_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<BF16, I32, BF16>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_maxpool_bwd_f32_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<F32, I32, F32>>>&);
+#endif
+template <typename DOutDataType, typename IndexDataType, typename DInDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>>
+{
+    using DeviceOp = DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<DOutDataType, F16> && is_same_v<DInDataType, F16> &&
+                     is_same_v<IndexDataType, I32>)
+            add_device_maxpool_bwd_f16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_BF16
+        else if constexpr(is_same_v<DOutDataType, BF16> && is_same_v<DInDataType, BF16> &&
+                          is_same_v<IndexDataType, I32>)
+            add_device_maxpool_bwd_bf16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_FP32
+        else if constexpr(is_same_v<DOutDataType, F32> && is_same_v<DInDataType, F32> &&
+                          is_same_v<IndexDataType, I32>)
+            add_device_maxpool_bwd_f32_instances(op_ptrs);
+#endif
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp

@@ -37,6 +37,21 @@ void add_device_pool3d_fwd_ndhwc_index_f16_instances(
     std::vector<std::unique_ptr<
         DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NDHWC, NDHWC, MaxOp, true>>>&);
 #endif
+#ifdef CK_ENABLE_BF16
+// BF16
+void add_device_pool3d_fwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NDHWC, NDHWC, MaxOp, false>>>&);
+
+void add_device_pool3d_fwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NDHWC, NDHWC, AvgOp, false>>>&);
+
+// BF16 - return index
+void add_device_pool3d_fwd_ndhwc_index_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NDHWC, NDHWC, MaxOp, true>>>&);
+#endif
 #ifdef CK_ENABLE_FP32
 // FP32
 void add_device_pool3d_fwd_ndhwc_f32_instances(
@@ -98,9 +113,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFw
                 }
             }
 #endif
+#ifdef CK_ENABLE_BF16
+            else if constexpr(is_same_v<InDataType, BF16> && is_same_v<OutDataType, BF16> &&
+                              is_same_v<IndexDataType, I32>)
+            {
+                if constexpr(OutputIndex && ReduceOpId == MaxOp)
+                {
+                    add_device_pool3d_fwd_ndhwc_index_bf16_instances(op_ptrs);
+                }
+                else
+                {
+                    add_device_pool3d_fwd_ndhwc_bf16_instances(op_ptrs);
+                }
+            }
+#endif
 #ifdef CK_ENABLE_FP32
-            if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
-                         is_same_v<IndexDataType, I32>)
+            else if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
+                              is_same_v<IndexDataType, I32>)
             {
                 if constexpr(OutputIndex && ReduceOpId == MaxOp)
                 {

library/include/ck/library/utility/check_err.hpp

@@ -65,7 +65,11 @@ check_err(const Range& out,
     }
     if(!res)
     {
-        std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
+        const float error_percent =
+            static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
+        std::cerr << "max err: " << max_err;
+        std::cerr << ", number of errors: " << err_count;
+        std::cerr << ", " << error_percent << "% wrong values" << std::endl;
     }
     return res;
 }
@@ -112,7 +116,11 @@ check_err(const Range& out,
     }
     if(!res)
     {
-        std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
+        const float error_percent =
+            static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
+        std::cerr << "max err: " << max_err;
+        std::cerr << ", number of errors: " << err_count;
+        std::cerr << ", " << error_percent << "% wrong values" << std::endl;
     }
     return res;
 }
@@ -158,7 +166,11 @@ check_err(const Range& out,
     }
     if(!res)
     {
-        std::cerr << std::setw(12) << std::setprecision(7) << "max err: " << max_err << std::endl;
+        const float error_percent =
+            static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
+        std::cerr << "max err: " << max_err;
+        std::cerr << ", number of errors: " << err_count;
+        std::cerr << ", " << error_percent << "% wrong values" << std::endl;
     }
     return res;
 }
@@ -209,7 +221,11 @@ check_err(const Range& out,
     }
     if(!res)
     {
-        std::cerr << "max err: " << max_err << std::endl;
+        const float error_percent =
+            static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
+        std::cerr << "max err: " << max_err;
+        std::cerr << ", number of errors: " << err_count;
+        std::cerr << ", " << error_percent << "% wrong values" << std::endl;
     }
     return res;
 }

library/include/ck/library/utility/device_memory.hpp

@@ -26,7 +26,9 @@ struct DeviceMem
     void* GetDeviceBuffer() const;
     std::size_t GetBufferSize() const;
     void ToDevice(const void* p) const;
+    void ToDevice(const void* p, const std::size_t cpySize) const;
     void FromDevice(void* p) const;
+    void FromDevice(void* p, const std::size_t cpySize) const;
     void SetZero() const;
     template <typename T>
     void SetValue(T x) const;

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt

+set(DEVICE_AVGPOOL_BWD_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_f16_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_f32_instance.cpp)
+endif()
+add_instance_library(device_avg_pool3d_bwd_instance ${DEVICE_AVGPOOL_BWD_INSTANCES})

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using I32   = int32_t;
+using F16   = ck::half_t;
+using BF16  = ck::bhalf_t;
+using F32   = float;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+using device_avgpool_bwd_ndhwc_f16_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+using device_avgpool_bwd_ndhwc_bf16_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+using device_avgpool_bwd_ndhwc_f32_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, BF16, BF16, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F16, F16, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F32, F32, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt

+add_instance_library(device_gemm_multiply_add_instance
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
+
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
+)

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp

@@ -83,7 +83,6 @@ using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_tile_instanc
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>
     // clang-format on
-    // clang-format on
     >;
 
 void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(