Merge branch 'develop' into conv_splitk_f32

composable_kernel/include/utility/common_header.hpp

@@ -35,8 +35,8 @@
 #include "dynamic_buffer.hpp"
 #include "is_known_at_compile_time.hpp"
 #include "transpose_vectors.hpp"
-
 #include "inner_product.hpp"
+#include "element_wise_operation.hpp"
 
 // TODO: remove this
 #if CK_USE_AMD_INLINE_ASM

composable_kernel/include/utility/config.hpp

@@ -24,12 +24,16 @@
 #define CK_MIN_BLOCK_PER_CU 2
 #endif
 
-// buffer resourse
+// GPU-specific parameters
 #if defined(CK_AMD_GPU_GFX803) || defined(CK_AMD_GPU_GFX900) || defined(CK_AMD_GPU_GFX906) || \
     defined(CK_AMD_GPU_GFX908) || defined(CK_AMD_GPU_GFX90A)
+// buffer resourse
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
+// wave size
+#define CK_GPU_WAVE_SIZE 64
 #elif defined(CK_AMD_GPU_GFX1030)
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
+#define CK_GPU_WAVE_SIZE 32
 #endif
 
 // FMA instruction
@@ -141,6 +145,10 @@
 #define CK_WORKAROUND_SWDEV_XXXXXX_THREAD_WISE_COPY_V1R4_TYPE_CONVERT_ISSUE 1
 #endif
 
+#ifndef CK_WORKAROUND_SWDEV_XXXXXX_THREAD_WISE_COPY_V1R5_TYPE_CONVERT_ISSUE
+#define CK_WORKAROUND_SWDEV_XXXXXX_THREAD_WISE_COPY_V1R5_TYPE_CONVERT_ISSUE 1
+#endif
+
 namespace ck {
 
 enum InMemoryDataOperationEnum_t
@@ -152,7 +160,7 @@ enum InMemoryDataOperationEnum_t
 
 enum ActivTypeEnum_t
 {
-    None = 0,
+    None,
     LeakyRelu,
     Sigmoid
 };