Home

Welcome to the kuda-12.2 wiki!

Purpose

The purpose of this library is to facilitate convenient usage of CUDA for Kotlin developers.

Version Management

Following Semantic Versioning (major.minor.patch)

Branch Rule

• develop : merge many githubid branches develop branch is the newest code (for versioning branch)
• githubid : it is personal-common branch for kudakit organization members

Common Coding Rule

1. Please Add function folling format

/**
* Function Description (cudaFunctionName)
*
* @param param Parameter Description
*
* @return Return Description
*/
external fun functionName(type : Type) : Type

for example,

/**
* Sets the event for an event wait node in the given graphExec. (cuGraphExecEventWaitNodeSetEvent)
*
* @param hGraphExec The executable graph in which to set the specified node
* @param hNode event wait node from the graph from which graphExec was instantiated
* @param event Updated event to use
*/
external fun graphExecEventWaitNodeSetEvent(hGraphExec : Long, hNode : Long, event : Long) : Int

2. Use Automatic Type Casting of JNI

for example 1, Here is code that needs cleaning up

JNIEXPORT jlong JNICALL Java_kuda_runtimeapi_DeviceHandler_getDefaultMemPool(JNIEnv* env, jclass cls, jint device) {

	cudaMemPool_t memPool;

	cudaError_t cudaStatus = cudaDeviceGetMemPool(&memPool, (int)device);

	if (cudaStatus != cudaSuccess) {
		return cudaStatus;
	}

	return (jlong)memPool;
}

remove type casting in 'same data type (jint<->int, jlong<->long, etc..)'

JNIEXPORT jlong JNICALL Java_kuda_runtimeapi_DeviceHandler_getDefaultMemPool(JNIEnv* env, jclass cls, jint device) {

	cudaMemPool_t memPool;

	cudaError_t cudaStatus = cudaDeviceGetMemPool(&memPool, device);

	if (cudaStatus != cudaSuccess) {
		return cudaStatus;
	}

	return (jlong)memPool;
}

example 2, before

JNIEXPORT jint JNICALL Java_kuda_kublas_Kublas_getPointerMode(JNIEnv* env, jobject obj, jlong handle) {

	//initialize...
	cublasPointerMode_t cublasPointerMode = CUBLAS_POINTER_MODE_HOST;

	cublasHandle_t cublasHandle = reinterpret_cast<cublasHandle_t>(handle);

	cublasStatus_t cublasStatus = cublasGetPointerMode(cublasHandle, &cublasPointerMode);

	if (cublasStatus != CUBLAS_STATUS_SUCCESS) {
		return cublasStatus;
	}

	return (jint)static_cast<int>(cublasPointerMode);
}

after

JNIEXPORT jint JNICALL Java_kuda_kublas_Kublas_getPointerMode(JNIEnv* env, jobject obj, jlong handle) {

	//initialize...
	cublasPointerMode_t cublasPointerMode = CUBLAS_POINTER_MODE_HOST;

	cublasHandle_t cublasHandle = reinterpret_cast<cublasHandle_t>(handle);

	cublasStatus_t cublasStatus = cublasGetPointerMode(cublasHandle, &cublasPointerMode);

	if (cublasStatus != CUBLAS_STATUS_SUCCESS) {
		return cublasStatus;
	}

	return static_cast<int>(cublasPointerMode);
}

3. Use Integer instead of Byte for Kotlin Enum Class

example. Lets write kotlin enum class for CUDA CUmemAllocationCompType enum

enum CUmemAllocationCompType : Specifies compression attribute for an allocation.

Values
- CU_MEM_ALLOCATION_COMP_NONE = 0x0
- CU_MEM_ALLOCATION_COMP_GENERIC = 0x1

Use a Integer value

enum class MemAllocationCompType(val num : Int) {
    NONE(0x0),
    GENERIC(0x1)
}

3. Keep Kotlin enum class forms

/**
 *
 * Descriptions (CUDA Enum name)
 *
 * @property PROPERTY Property Descriptions
 */
enum class IpcMemFlags(val num : Int) {
    PROPERTY(0)
}

for example,

/**
 *
 * CUDA Ipc Mem Flags (CUipcMem_flags)
 *
 * @property LAZY_ENABLE_PEER_ACCESS Automatically enable peer access between remote devices as needed
 */
enum class IpcMemFlags(val num : Int) {
    LAZY_ENABLE_PEER_ACCESS(0x1)
}