About green context in cuda13.2.1

Accelerated Computing CUDA CUDA Programming and Performance
1

I want to try the green context using cuda 13.2.1. but I find some apis can not find and can not compile. such as cudaDevResourceDescDestroy and cudaGreenCtxDestroy, is the two api have not included in cuda 13.2.1?

2

I want to do some experiments on the Blackwell GPU and I’m not sure if CUDA 13.2 is support green context

3

Can you state which operating system you used? Which compiler? Is it a compiler or linker error?

4

My os is win10 x64, this is a compile error:

error : identifier “cudaGreenCtxDestroy” is undefined.

By the way, the following code run errror: cuda error code is 915, the gpu this blackwell PRO 4000

int main() {
	// 1. 初始化并设置设备
	int deviceId = 0;
	CUDA_CHECK(cudaSetDevice(deviceId));

	cudaDeviceProp deviceProp = { 0 };
	cudaGetDeviceProperties(&deviceProp, deviceId);

	// 2. 获取设备的全部 SM 资源信息
	cudaDevResource fullResource;
	CUDA_CHECK(cudaDeviceGetDevResource(deviceId, &fullResource, cudaDevResourceTypeSm));

	unsigned int totalSMs = fullResource.sm.smCount;
	printf("Total SMs on device: %u, %s\n", totalSMs, deviceProp.name);
	if (totalSMs < 24) {
		fprintf(stderr, "Error: Device has only %u SMs, need at least 24.\n", totalSMs);
		return -1;
	}

	// 3. 拆分 SM 资源 (使用 SplitByCount)
	//    目标: 创建两个分区，分别包含 16 个和 8 个 SM
	unsigned int smCounts[] = { 16, 8 };
	const unsigned int numPartitions = 2;

	// 定义数组来存储拆分结果和剩余资源
	cudaDevResource partitions[numPartitions];
	cudaDevResource remainingResource; // <- 关键点入下

	for (unsigned int i = 0; i < numPartitions; ++i) {
		// 注意：这里才是真正的单次拆分，每次拆分都需要更新输入
		// 如果用 baseResource = fullResource 反复传入，逻辑是错的！
		cudaDevResource inputResource = (i == 0) ? fullResource : remainingResource;
		unsigned int nbGroups = 1; // 每次请求创建一个分区

		CUDA_CHECK(cudaDevSmResourceSplitByCount(
			&partitions[i], // 输出: 本次拆分到的分区资源
			&nbGroups,              // 输出: 实际创建的分区数，应为 1 (成功时)
			&inputResource,         // 输入: 用于拆分的源资源
			&remainingResource,     // 输出: 拆分后剩余的源资源
			0,                      // flags (保留为0)
			smCounts[i]             // 输入: 请求的 SM 数量，API 会自动处理对齐
		));
		printf("Partition %u requested %u SMs.\n", i, smCounts[i]);
	}

	// 5. 为每个分区创建资源描述符，然后分别创建Green Context
	cudaExecutionContext_t greenCtx[2] = { nullptr, nullptr };
	for (int i = 0; i < 2; ++i) {
		cudaDevResourceDesc_t resourceDesc = nullptr;
		CUDA_CHECK(cudaDevResourceGenerateDesc(&resourceDesc, &partitions[i], 1));
		CUDA_CHECK(cudaGreenCtxCreate(&greenCtx[i], resourceDesc, deviceId, 0));
		//		CUDA_CHECK(cudaDevResourceDescDestroy(resourceDesc));
		std::cout << "Green Context " << i << " created with " << partitions[i].sm.smCount << " SMs." << std::endl;
	}

	// 6. 在各自的Green Context上创建流
	cudaStream_t streams[2];
	for (int i = 0; i < 2; ++i) {
		CUDA_CHECK(cudaExecutionCtxStreamCreate(&streams[i], greenCtx[i], cudaStreamDefault, 0));
	}



	printf("Green Contexts created successfully.\n");

}
5

There is only cuGreenCtxDestroy, not cuda. That means only a driver API function.
I believe you cannot destroy green contexts without the driver API.

6

in cuda 13.2, it has cudaGreenCtxCreate but has no cudaGreenCtxDestroy function?

my green context is created by cudaGreenCtxCreate, so it can not destroyed by the driver api

7

The function to destroy green context is

__host__ cudaError_t CUDARTAPI cudaExecutionCtxDestroy(cudaExecutionContext_t ctx)
8

thank you.

Also, do I need to destroy resourceDesc in the following code snippet? If necessary, which function should be called.

// 5. 为每个分区创建资源描述符，然后分别创建Green Context
cudaExecutionContext_t greenCtx[2] = { nullptr, nullptr };
for (int i = 0; i < 2; ++i) {
	cudaDevResourceDesc_t resourceDesc = nullptr;
	CUDA_CHECK(cudaDevResourceGenerateDesc(&resourceDesc, &partitions[i], 1));
	CUDA_CHECK(cudaGreenCtxCreate(&greenCtx[i], resourceDesc, deviceId, 0));
//	CUDA_CHECK(cudaDevResourceDescDestroy(resourceDesc));
	std::cout << "Green Context " << i << " created with " << partitions[i].sm.smCount << " SMs." << std::endl;
}
9

I do not know. I have not used green context before.

If you cannot find this information in the programming guide 4.6. Green Contexts — CUDA Programming Guide or runtime API CUDA Runtime API :: CUDA Toolkit Documentation you could file a bug asking for clarification

10

CUDA_CHECK(cudaDevSmResourceSplitByCount(
&partitions[i], // 输出: 本次拆分到的分区资源
&nbGroups, // 输出: 实际创建的分区数，应为 1 (成功时)
&inputResource, // 输入: 用于拆分的源资源
&remainingResource, // 输出: 拆分后剩余的源资源
0, // flags (保留为0)
smCounts[i] // 输入: 请求的 SM 数量，API 会自动处理对齐
));

the function return 915,

/**
* This error indicates one or more resources are insufficient or non-applicable for
* the operation.
*/
cudaErrorInvalidResourceConfiguration = 915,

Is there anyone who can give some advice？

Thanks