CUDNN_STATUS_NOT_SUPPORTED when finalizing engine config for simple ReLU graph

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

I’m roughly following Setting Up An Operation Graph For A Grouped Convolution in order to setup a simple ReLU forward pass but I’m getting CUDNN_STATUS_NOT_SUPPORTED when trying to finalize the engine config.

Here’s my code:

#include <cuda_runtime.h>
#include <cudnn.h>
#include <iostream>
#include <cstdio>

#define CHECK_CUDNN(expression) \
{ \
    cudnnStatus_t status = (expression); \
    if(status != CUDNN_STATUS_SUCCESS) \
    { \
        std::cerr << "Error on line " << __LINE__ << ": " \
            << cudnnGetErrorString(status) << std::endl; \
        std::exit(EXIT_FAILURE); \
    } \
}

void tensor_2d_create(
    int64_t dim0, int64_t dim1, int64_t* tensor_count, cudnnBackendDescriptor_t* desc
){
    int64_t n_dims = 2;
    int64_t shape[] = {dim0, dim1};
    int64_t strides[] = {dim1, 1};
    int64_t alignment = 16;
    int64_t uid = (*tensor_count)++;;

    cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
    CHECK_CUDNN(cudnnBackendCreateDescriptor(CUDNN_BACKEND_TENSOR_DESCRIPTOR, desc));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        *desc, CUDNN_ATTR_TENSOR_DATA_TYPE, CUDNN_TYPE_DATA_TYPE, 1, &data_type
    ));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        *desc, CUDNN_ATTR_TENSOR_DIMENSIONS, CUDNN_TYPE_INT64, n_dims, shape
    ));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        *desc, CUDNN_ATTR_TENSOR_STRIDES, CUDNN_TYPE_INT64, n_dims, strides
    ));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        *desc, CUDNN_ATTR_TENSOR_BYTE_ALIGNMENT, CUDNN_TYPE_INT64, 1, &alignment
    ));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        *desc, CUDNN_ATTR_TENSOR_UNIQUE_ID, CUDNN_TYPE_INT64, 1, &uid
    ));
    CHECK_CUDNN(cudnnBackendFinalize(*desc));
}

int main()
{
    cudnnHandle_t cudnn;
    CHECK_CUDNN(cudnnCreate(&cudnn));
    printf("Initialized cuDNN\n");
    printf("cuDNN version: %zu\n", cudnnGetVersion());

    int64_t tensor_count = 0;
    cudnnBackendDescriptor_t input_desc;
    tensor_2d_create(1, 32, &tensor_count, &input_desc);
    
    cudnnBackendDescriptor_t output_desc;
    tensor_2d_create(1, 32, &tensor_count, &output_desc);

    cudnnPointwiseMode_t act_mode = CUDNN_POINTWISE_RELU_FWD;
    cudnnDataType_t act_data_type = CUDNN_DATA_FLOAT;
    cudnnBackendDescriptor_t relu_desc;
    CHECK_CUDNN(cudnnBackendCreateDescriptor(CUDNN_BACKEND_POINTWISE_DESCRIPTOR, &relu_desc));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        relu_desc, CUDNN_ATTR_POINTWISE_MODE, CUDNN_TYPE_POINTWISE_MODE, 1, &act_mode
    ));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        relu_desc, CUDNN_ATTR_POINTWISE_MATH_PREC, CUDNN_TYPE_DATA_TYPE, 1, &act_data_type
    ));
    CHECK_CUDNN(cudnnBackendFinalize(relu_desc));

    cudnnBackendDescriptor_t relu_op_desc;
    CHECK_CUDNN(cudnnBackendCreateDescriptor(CUDNN_BACKEND_OPERATION_POINTWISE_DESCRIPTOR, &relu_op_desc));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        relu_op_desc, CUDNN_ATTR_OPERATION_POINTWISE_PW_DESCRIPTOR, CUDNN_TYPE_BACKEND_DESCRIPTOR, 1, &relu_desc
    ));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        relu_op_desc, CUDNN_ATTR_OPERATION_POINTWISE_XDESC, CUDNN_TYPE_BACKEND_DESCRIPTOR, 1, &input_desc
    ));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        relu_op_desc, CUDNN_ATTR_OPERATION_POINTWISE_YDESC, CUDNN_TYPE_BACKEND_DESCRIPTOR, 1, &output_desc
    ));
    CHECK_CUDNN(cudnnBackendFinalize(relu_op_desc));
    printf("Final tensor_count: %ld\n", tensor_count);

    // Create op graph.
    cudnnBackendDescriptor_t op_graph;
    CHECK_CUDNN(cudnnBackendCreateDescriptor(CUDNN_BACKEND_OPERATIONGRAPH_DESCRIPTOR, &op_graph));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        op_graph, CUDNN_ATTR_OPERATIONGRAPH_OPS, CUDNN_TYPE_BACKEND_DESCRIPTOR, 1, &relu_op_desc
    ));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        op_graph, CUDNN_ATTR_OPERATIONGRAPH_HANDLE, CUDNN_TYPE_HANDLE, 1, &cudnn
    ));
    CHECK_CUDNN(cudnnBackendFinalize(op_graph));
    printf("Created graph\n");

    // Create engine.
    cudnnBackendDescriptor_t engine;
    CHECK_CUDNN(cudnnBackendCreateDescriptor(CUDNN_BACKEND_ENGINE_DESCRIPTOR, &engine));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        engine, CUDNN_ATTR_ENGINE_OPERATION_GRAPH, CUDNN_TYPE_BACKEND_DESCRIPTOR, 1, &op_graph
    ));
    int64_t gidx = 0;
    CHECK_CUDNN(cudnnBackendSetAttribute(
        engine, CUDNN_ATTR_ENGINE_GLOBAL_INDEX, CUDNN_TYPE_INT64, 1, &gidx
    ));
    CHECK_CUDNN(cudnnBackendFinalize(engine));
    printf("Created engine\n");

    // Create engine config.
    cudnnBackendDescriptor_t engine_cfg;
    CHECK_CUDNN(cudnnBackendCreateDescriptor(CUDNN_BACKEND_ENGINECFG_DESCRIPTOR, &engine_cfg));
    CHECK_CUDNN(cudnnBackendSetAttribute(
        engine_cfg, CUDNN_ATTR_ENGINECFG_ENGINE, CUDNN_TYPE_BACKEND_DESCRIPTOR, 1, &engine
    ));
    // Error here.
    CHECK_CUDNN(cudnnBackendFinalize(engine_cfg));
    int64_t workspace_size;
    CHECK_CUDNN(cudnnBackendGetAttribute(
        engine_cfg, CUDNN_ATTR_ENGINECFG_WORKSPACE_SIZE, CUDNN_TYPE_INT64, 1, NULL, &workspace_size
    ));
    printf("Created engine config\n");

    CHECK_CUDNN(cudnnDestroy(cudnn));
}

cmake_minimum_required(VERSION 3.10)
project(mlp CUDA CXX)

set(CMAKE_CXX_STANDARD 14)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CUDA_STANDARD 14)
set(CMAKE_CUDA_STANDARD_REQUIRED ON)
set(CMAKE_CUDA_ARCHITECTURES "86")

find_package(CUDA REQUIRED)

set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-O3;-arch=sm_86)

cuda_add_executable(mlp_cudnn_iso ./src/mlp_cudnn_iso.cu)
set_target_properties(mlp_cudnn_iso PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
target_include_directories(mlp_cudnn_iso PRIVATE ./src/)
target_link_libraries(mlp_cudnn_iso cuda cudart cudnn cublas)

And here’s the logs with CUDNN_LOGLEVEL_DBG=2:

Initialized cuDNN
cuDNN version: 90300
Final tensor_count: 2
Created graph
Created engine

W! CuDNN (v90300 75) function cudnnBackendFinalize() called:
w!         Warning: CUDNN_STATUS_NOT_SUPPORTED_ARCH_MISMATCH; Reason: fortBackend.is_cutlass_sm7x() && ((this->getDeviceProp()->deviceVer < 700) || (this->getDeviceProp()->deviceVer > 800))
w!         Warning: CUDNN_STATUS_NOT_SUPPORTED_ARCH_MISMATCH; Reason: init_kernelgen_backend()
w!         Warning: CUDNN_STATUS_NOT_SUPPORTED; Reason: ptr->isSupported()
w!         Warning: CUDNN_STATUS_NOT_SUPPORTED; Reason: finalize_internal()
w!         Warning: CUDNN_STATUS_NOT_SUPPORTED; Reason: ptrDesc->finalize()
w! Time: 2024-11-01T13:17:40.153604 (0d+0h+0m+1s since start)
w! Process=3161219; Thread=3161219; GPU=NULL; Handle=NULL; StreamId=NULL.

Error on line 117: CUDNN_STATUS_NOT_SUPPORTED

My CUDA version is 12.4, my cuDNN version is 9.3, and I’m running this on a 3070 Ti.

The logs seem to indicate that the problem is related to the target architecture but I’m not sure how to interpret it. Any help would be appreciated.

2

Hi ,
Checking this with the Engineering and will update the thread.
Thanks

3

Hi,
Try 3D Tensor instead of 2D Tensors.

Maybe check the documentation on Support Surface and the constrains for certain Nodes/Operations: Graph API — NVIDIA cuDNN