Merge branch 'develop' into jd/dev_pkg

dd6a8de4 · Jehandad Khan · 0aa899aa · abf4bdb9 · dd6a8de4 · dd6a8de4
Commit dd6a8de4 authored Apr 06, 2022 by Jehandad Khan
20 changed files
--- a/include/ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_bias.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_bias.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle.hpp
@@ -138,7 +138,7 @@ struct DeviceGemmXdl_C_Shuffle
        AccDataType,
        CShuffleDataType,
        CDataType,
-        InMemoryDataOperationEnum_t::Set,
+        InMemoryDataOperationEnum::Set,
        AGridDesc_K0_M_K1,
        BGridDesc_K0_N_K1,
        CGridDesc_M_N,

--- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_2d.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_2d.hpp
@@ -4,9 +4,7 @@
 #include <iostream>
 #include <sstream>
 #include "device.hpp"
-#include "device_base.hpp"
-#include "device_gemm.hpp"
-#include "device_gemm_xdl.hpp"
+#include "device_gemm_bias.hpp"
 #include "common_header.hpp"
 #include "tensor_layout.hpp"
 #include "tensor_descriptor.hpp"
@@ -141,7 +139,7 @@ struct DeviceGemmXdl_C_Shuffle_Bias_2d
        ADataType, // TODO: distinguish A/B datatype
        AccDataType,
        CDataType,
-        InMemoryDataOperationEnum_t::Set,
+        InMemoryDataOperationEnum::Set,
        AGridDesc_K0_M_K1,
        BGridDesc_K0_N_K1,
        CGridDesc_M_N,

--- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_activation.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_activation.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_activation_add.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_activation_add.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_cshuffle.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_cshuffle.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk.hpp
--- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp
--- a/include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp
--- a/include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp
--- a/include/ck/tensor_operation/gpu/device/device_pool2d_fwd_nhwc_nhwc.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_pool2d_fwd_nhwc_nhwc.hpp
--- a/include/ck/tensor_operation/gpu/device/device_reduce.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_reduce.hpp
--- a/include/ck/tensor_operation/gpu/device/device_reduce_blockwise.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_reduce_blockwise.hpp
--- a/include/ck/tensor_operation/gpu/device/device_reduce_blockwise_second_call.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_reduce_blockwise_second_call.hpp
--- a/include/ck/tensor_operation/gpu/device/device_reduce_common.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_reduce_common.hpp