Merge branch 'develop' into feature/integrage-karg-simplification-pr

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp

@@ -50,8 +50,9 @@ __global__ void
             const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
-    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||             \
+    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
+    defined(__gfx1101__) || defined(__gfx1102__))
 
     constexpr index_t shared_block_size =
         GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
@@ -553,7 +554,8 @@ struct DeviceGemmMultipleD_Dl : public DeviceGemmMultipleD<ALayout,
     {
         if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" ||
            ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx1030" ||
-           ck::get_device_name() == "gfx940")
+           ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" ||
+           ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102")
         {
             return GridwiseGemm::CheckValidity(
                 arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_);

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp

@@ -1027,7 +1027,9 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     static bool IsSupportedArgument(const Argument& arg)
     {
         // check device
-        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
+             ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+             ck::get_device_name() == "gfx1102"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp

@@ -39,8 +39,9 @@ __global__ void
                                           const BElementwiseOperation b_element_op,
                                           const CDEElementwiseOperation cde_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
-    defined(__gfx90a__) || defined(__gfx1030__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||              \
+    defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__) || defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();

include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp

@@ -28,6 +28,7 @@ template <typename InDataType,
           typename AccElementwiseOperation,
           bool PropagateNan,
           bool OutputIndex,
+          bool TransformIndexKtoGlobal,
           bool HaveIndexInputIfOutputIndex,
           index_t BlockSize,
           index_t MThreadSliceSize,
@@ -260,6 +261,7 @@ struct DeviceReduceThreadWise : public DeviceReduce<InDataType,
 
             const auto kernel = kernel_reduce_threadwise<GridwiseReduce,
                                                          OutputIndex,
+                                                         TransformIndexKtoGlobal,
                                                          HaveIndexInput,
                                                          InDataType,
                                                          OutDataType,

include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp

@@ -15,6 +15,7 @@ namespace ck {
 
 template <typename GridwiseReduction,
           bool OutputIndex,
+          bool TransformIndexKtoGlobal,
           bool HaveIndexInput,
           typename InDataType,
           typename OutDataType,
@@ -48,16 +49,17 @@ __global__ void kernel_reduce_threadwise(const InGridDesc_M_K in_grid_desc_m_k,
     }
     else
     {
-        GridwiseReduction::template RunWithIndex<HaveIndexInput>(in_grid_desc_m_k,
-                                                                 out_grid_desc_m,
-                                                                 in_elementwise_op,
-                                                                 acc_elementwise_op,
-                                                                 alpha,
-                                                                 p_in_value_global,
-                                                                 p_in_index_global,
-                                                                 beta,
-                                                                 p_out_value_global,
-                                                                 p_out_index_global);
+        GridwiseReduction::template RunWithIndex<TransformIndexKtoGlobal, HaveIndexInput>(
+            in_grid_desc_m_k,
+            out_grid_desc_m,
+            in_elementwise_op,
+            acc_elementwise_op,
+            alpha,
+            p_in_value_global,
+            p_in_index_global,
+            beta,
+            p_out_value_global,
+            p_out_index_global);
     };
 };
 
@@ -232,7 +234,7 @@ struct GridwiseReduction_mk_to_m_threadwise
             reduced_data_desc, make_tuple(I0), accu_value_buf, out_grid_desc_m, dst_global_buf);
     };
 
-    template <bool HaveIndexInput>
+    template <bool TransformIndexKtoGlobal, bool HaveIndexInput>
     __device__ static void RunWithIndex(const InGridDesc_M_K& in_grid_desc_m_k,
                                         const OutGridDesc_M& out_grid_desc_m,
                                         const InElementwiseOperation& in_elementwise_op,
@@ -390,6 +392,18 @@ struct GridwiseReduction_mk_to_m_threadwise
                 indexStart += KThreadSliceSize;
                 reducedLength += KThreadSliceSize;
             } while(reducedLength < toReduceLength);
+
+            if constexpr(TransformIndexKtoGlobal)
+            {
+                static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                    const auto coord = make_tensor_coordinate(
+                        in_grid_desc_m_k,
+                        make_multi_index(thread_global_1d_id * MThreadSliceSize + I,
+                                         accu_index_buf(I)));
+
+                    accu_index_buf(I) = coord.GetOffset();
+                });
+            }
         };
 
         // for indiced operation, acc_elementwise_op shoud do nothing

include/ck/utility/amd_llvm_intrinsic.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_AMD_LLVM_INTRINSIC_HPP
-#define CK_AMD_LLVM_INTRINSIC_HPP
-
-#include "data_type.hpp"
-
-namespace ck {
-
-__device__ int32_t llvm_amdgcn_readfirstlane_i32(int32_t i) __asm("llvm.amdgcn.readfirstlane");
-
-} // namespace ck
-#endif

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

@@ -3,8 +3,8 @@
 
 #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_gemm.hpp"