Merge remote-tracking branch 'origin/develop' into batched_gemm_c_permute

library/CMakeLists.txt

-add_subdirectory(src/host_tensor)
 add_subdirectory(src/tensor_operation_instance/gpu)
+add_subdirectory(src/host_tensor)
 add_subdirectory(src/utility)

library/include/ck/library/host_tensor/conv_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include "ck/tensor_description/tensor_descriptor.hpp"

library/include/ck/library/host_tensor/device_memory.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <hip/hip_runtime.h>

library/include/ck/library/host_tensor/host_common_util.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <vector>

library/include/ck/library/host_tensor/host_conv.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 #include "host_tensor.hpp"
 #include "conv_common.hpp"

library/include/ck/library/host_tensor/host_gemm.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include "host_tensor.hpp"

library/include/ck/library/host_tensor/host_reduction.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <vector>

library/include/ck/library/host_tensor/host_tensor.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <thread>
@@ -219,6 +222,12 @@ struct Tensor
 
     Tensor(const Tensor& other) : mDesc(other.mDesc), mData(other.mData) {}
 
+    Tensor& operator=(const Tensor& other)
+    {
+        mDesc = other.mDesc;
+        mData = other.mData;
+    }
+
     template <typename F>
     void ForEach_impl(F&& f, std::vector<size_t>& idx, size_t rank)
     {
@@ -361,13 +370,8 @@ HostTensorDescriptor::HostTensorDescriptor(const std::vector<X>& lens,
 {
 }
 
-void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream& os = std::cout);
-
 #if 1
 // FIXME: remove
-void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst);
-#endif
-
 template <typename T>
 float check_error(const Tensor<T>& ref, const Tensor<T>& result)
 {
@@ -413,3 +417,4 @@ float check_error(const Tensor<T>& ref, const Tensor<T>& result)
 
     return linf_error;
 }
+#endif

library/include/ck/library/host_tensor/host_tensor_generator.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <cmath>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>
@@ -59,20 +62,20 @@ struct ReferenceBatchedGemm : public device::BaseOperator
 
                 for(int k = 0; k < K; ++k)
                 {
-                    float v_a;
-                    float v_b;
+                    ADataType v_a;
+                    BDataType v_b;
 
-                    arg.a_element_op_(v_a, static_cast<const float>(arg.a_g_m_k_(g, m, k)));
-                    arg.b_element_op_(v_b, static_cast<const float>(arg.b_g_k_n_(g, k, n)));
+                    arg.a_element_op_(v_a, arg.a_g_m_k_(g, m, k));
+                    arg.b_element_op_(v_b, arg.b_g_k_n_(g, k, n));
 
-                    v_acc += v_a * v_b;
+                    v_acc += ck::type_convert<float>(v_a) * ck::type_convert<float>(v_b);
                 }
 
                 float v_c;
 
                 arg.c_element_op_(v_c, v_acc);
 
-                arg.c_g_m_n_(g, m, n) = v_c;
+                arg.c_g_m_n_(g, m, n) = ck::type_convert<CDataType>(v_c);
             };
 
             make_ParallelTensorFunctor(f_gmk_gkn_gmn,

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>
@@ -60,20 +63,21 @@ struct ReferenceGemm : public device::BaseOperator
 
                 for(int k = 0; k < K; ++k)
                 {
-                    AccDataType v_a;
-                    AccDataType v_b;
+                    ADataType v_a;
+                    BDataType v_b;
 
-                    arg.a_element_op_(v_a, static_cast<const AccDataType>(arg.a_m_k_(m, k)));
-                    arg.b_element_op_(v_b, static_cast<const AccDataType>(arg.b_k_n_(k, n)));
+                    arg.a_element_op_(v_a, arg.a_m_k_(m, k));
+                    arg.b_element_op_(v_b, arg.b_k_n_(k, n));
 
-                    v_acc += v_a * v_b;
+                    v_acc +=
+                        ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
                 }
 
                 AccDataType v_c;
 
                 arg.c_element_op_(v_c, v_acc);
 
-                arg.c_m_n_(m, n) = v_c;
+                arg.c_m_n_(m, n) = ck::type_convert<CDataType>(v_c);
             };
 
             make_ParallelTensorFunctor(

library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 
 #include <iostream>