compile ck for all targets

README.md

@@ -20,7 +20,7 @@ mkdir build && cd build
 cmake                                                                 \
 -D BUILD_DEV=OFF                                                      \
 -D CMAKE_BUILD_TYPE=Release                                           \
--D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3  \
+-D CMAKE_CXX_FLAGS=" -O3  \
 -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                             \
 -D CMAKE_PREFIX_PATH=/opt/rocm                                        \
 ..

include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp

@@ -54,6 +54,7 @@ __global__ void
             const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
             const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -933,6 +934,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
 
         return str.str();
     }
+#endif //end of if defined (defined(__gfx908__) || defined(__gfx90a__))
 };
 
 } // namespace device

include/ck/tensor_operation/gpu/device/device_batched_gemm_xdl.hpp

@@ -46,6 +46,7 @@ __global__ void
             const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
             const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -609,6 +610,7 @@ struct DeviceBatchedGemmXdl
 
         return str.str();
     }
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 };
 
 } // namespace device

include/ck/tensor_operation/gpu/device/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp

@@ -49,6 +49,7 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / num_batches);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -654,6 +655,7 @@ struct DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_
 
         return str.str();
     }
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 };
 
 } // namespace device

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

@@ -55,7 +55,7 @@ __global__ void kernel_reduce_threadwise(const InGridDesc_M_K in_grid_desc_m_k,
                                          OutDataType* const __restrict__ p_out_global,
                                          IndexDataType* const __restrict__ p_indices_global)
 {
-    if constexpr(!NeedIndices)
+   if constexpr(!NeedIndices)
     {
         GridwiseReduction::Run(in_grid_desc_m_k,
                                out_grid_desc_m,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp

@@ -48,6 +48,7 @@ __global__ void
             const DGridDescriptor_MBlock_MPerBlock d_grid_desc_mblock_mperblock,
             const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
@@ -66,6 +67,7 @@ __global__ void
                                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
                                                    d_grid_desc_mblock_mperblock,
                                                    block_2_ctile_map);
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <typename FloatAB,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp

@@ -38,6 +38,7 @@ __global__ void
                                         c_grid_desc_mblock_mperblock_nblock_nperblock,
                                     const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
@@ -51,6 +52,7 @@ __global__ void
                                                    b_grid_desc_bk0_n_bk1,
                                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
                                                    block_2_ctile_map);
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <typename FloatAB,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp

@@ -39,6 +39,7 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
@@ -52,6 +53,7 @@ __global__ void
                                                    b_element_op,
                                                    c_element_op,
                                                    block_2_ctile_map);
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <typename GridwiseGemm,
@@ -74,6 +76,7 @@ __global__ void
             const BElementwiseOperation b_element_op,
             const CElementwiseOperation c_element_op)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();
@@ -126,6 +129,7 @@ __global__ void
         gemm_desc_ptr[group_id].block_2_ctile_map_,
         block_id_grp);
 #endif
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <index_t BlockSize,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp

@@ -37,6 +37,7 @@ __global__ void
                                 const CElementwiseOperation c_element_op,
                                 const CBlockClusterAdaptor c_block_cluster_adaptor)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     constexpr index_t shared_block_size =
         GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
 
@@ -53,6 +54,7 @@ __global__ void
                                                   b_element_op,
                                                   c_element_op,
                                                   c_block_cluster_adaptor);
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <index_t BlockSize,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp

@@ -39,6 +39,7 @@ __global__ void
                                   const CElementwiseOperation c_element_op,
                                   const CBlockClusterAdaptor c_block_cluster_adaptor)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     constexpr index_t shared_block_size =
         GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
 
@@ -55,6 +56,7 @@ __global__ void
                                                   b_element_op,
                                                   c_element_op,
                                                   c_block_cluster_adaptor);
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <index_t BlockSize,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp

@@ -42,6 +42,7 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainK0BlockLoop>(
@@ -56,6 +57,7 @@ __global__ void
         b_element_op,
         c_element_op,
         block_2_ctile_map);
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp

@@ -45,6 +45,7 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainK0BlockLoop>(
@@ -61,6 +62,7 @@ __global__ void
         b_element_op,
         c_element_op,
         block_2_ctile_map);
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp

@@ -49,6 +49,7 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
+#if (defined(__gfx908__) || defined(__gfx90a__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainK0BlockLoop>(
@@ -67,6 +68,7 @@ __global__ void
         b_element_op,
         c_element_op,
         block_2_ctile_map);
+#endif //end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <

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

@@ -36,6 +36,7 @@ __global__ void kernel_buffer_set_value(const Grid1dBufferDescType grid_1d_buffe
                                         DataType value)
 
 {
+
     using PassThroughOp = tensor_operation::element_wise::UnaryIdentic<DataType, DataType>;
 
     constexpr auto I0 = Number<0>{};

script/cmake-rocm.sh

@@ -10,7 +10,7 @@ cmake
 -D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL}                                                                                                  \
 -D BUILD_DEV=OFF                                                                                                                               \
 -D CMAKE_BUILD_TYPE=Release                                                                                                                    \
--D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD"         \
+-D CMAKE_CXX_FLAGS=" -O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD"         \
 -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                                                                      \
 -D CMAKE_PREFIX_PATH=/opt/rocm                                                                                                                 \
 -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                                                              \