bring ggml v2_cuda up to date with AMD changes

Author: YellowRoseCx

otherarch/ggml_v2-cuda-legacy.cu

@@ -8,6 +8,10 @@
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif
 #define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
@@ -32,6 +36,7 @@
 #define cudaEventDisableTiming hipEventDisableTiming
 #define cudaEventRecord hipEventRecord
 #define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
 #define cudaFree hipFree
 #define cudaFreeHost hipHostFree
 #define cudaGetDevice hipGetDevice
@@ -54,7 +59,7 @@
 #define cudaStreamCreateWithFlags hipStreamCreateWithFlags
 #define cudaStreamNonBlocking hipStreamNonBlocking
 #define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStreamWaitEvent hipStreamWaitEvent
+#define cudaStreamWaitEvent(stream, event) hipStreamWaitEvent(stream, event, 0)
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
 #else

otherarch/ggml_v2-cuda.cu

@@ -8,6 +8,10 @@
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif
 #define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
@@ -32,6 +36,7 @@
 #define cudaEventDisableTiming hipEventDisableTiming
 #define cudaEventRecord hipEventRecord
 #define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
 #define cudaFree hipFree
 #define cudaFreeHost hipHostFree
 #define cudaGetDevice hipGetDevice
@@ -54,14 +59,13 @@
 #define cudaStreamCreateWithFlags hipStreamCreateWithFlags
 #define cudaStreamNonBlocking hipStreamNonBlocking
 #define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStreamWaitEvent hipStreamWaitEvent
+#define cudaStreamWaitEvent(stream, event) hipStreamWaitEvent(stream, event, 0)
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
 #else
 #include <cuda_runtime.h>
 #include <cublas_v2.h>
 #include <cuda_fp16.h>
-
 #endif
 
 #include "ggml_v2-cuda.h"