[gfx110x] support Navi3x architectures. (#628)

* enable building on Nav31

* fix syntax

* replace GPU_TARGETS with offload-arch

* add gfx1102 rachitecture

* fix typo

* update changelog

[ROCm/composable_kernel commit: 0ccecc7c31]

include/ck/ck.hpp

@@ -27,14 +27,6 @@
 #define CK_WAVELET_MIN_BLOCK_PER_CU 2
 #endif
 
-// check GPU target
-#ifdef __HIP_DEVICE_COMPILE__
-#if !(defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
-      defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__))
-#error Not supported target
-#endif
-#endif
-
 // buffer resource
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_BUFFER_RESOURCE_3RD_DWORD -1
@@ -43,7 +35,7 @@
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
 #elif defined(__gfx1030__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
-#elif defined(__gfx1100__) // for GPU code
+#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x10020000
 #endif
 
@@ -72,7 +64,7 @@
 // WMMA instruction
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_WMMA
-#elif defined(__gfx1100__) // for GPU code
+#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
 #define CK_USE_AMD_WMMA
 #endif
 

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

@@ -770,7 +770,8 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::get_device_name() == "gfx1100")
+        if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -476,7 +476,8 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::get_device_name() == "gfx1100")
+        if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -404,7 +404,8 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::get_device_name() == "gfx1100")
+        if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -579,7 +579,8 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
         namespace ctc = tensor_layout::convolution;
 
         // check device
-        if(get_device_name() == "gfx1100")
+        if(get_device_name() == "gfx1100" || get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -54,7 +54,8 @@ __global__ void
             const Block2CTileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -147,7 +148,8 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const Block2CTileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
     // printf("entry kernel launch");
     __shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
 
@@ -242,7 +244,8 @@ __global__ void
             const CDEElementwiseOperation cde_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
     __shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
 
     GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -271,7 +274,7 @@ __global__ void
     ignore = b_element_op;
     ignore = cde_element_op;
     ignore = block_2_ctile_map;
-#endif // end of if (defined(__gfx1100__))
+#endif // end of if (defined(__gfx1100__ ))
 }
 
 template < // DataType Family

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

@@ -49,7 +49,8 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,