Porting CUDA driver API

Porting CUDA driver API#

NVIDIA provides separate CUDA driver and runtime APIs. The two APIs have significant overlap in functionality:

  • Both APIs support events, streams, memory management, memory copy, and error handling.

  • Both APIs deliver similar performance.

  • Driver API calls begin with the prefix cu, while runtime API calls begin with the prefix cuda. For example, the driver API contains cuEventCreate, while the runtime API contains cudaEventCreate, which has similar functionality.

  • The driver API defines a different, but largely overlapping, error code space than the runtime API and uses a different coding convention. For example, the driver API defines CUDA_ERROR_INVALID_VALUE, while the runtime API defines cudaErrorInvalidValue.

The driver API offers two additional functionalities not provided by the runtime API: cuModule and cuCtx APIs.

cuModule API#

The Module section of the driver API provides additional control over how and when accelerator code objects are loaded. For example, the driver API enables code objects to load from files or memory pointers. Symbols for kernels or global data are extracted from the loaded code objects. In contrast, the runtime API loads automatically and, if necessary, compiles all the kernels from an executable binary when it runs. In this mode, kernel code must be compiled using NVCC so that automatic loading can function correctly.

The Module features are useful in an environment that generates the code objects directly, such as a new accelerator language front end. NVCC is not used here. Instead, the environment might have a different kernel language or compilation flow. Other environments have many kernels and don’t want all of them to be loaded automatically. The Module functions load the generated code objects and launch kernels. Similar to the cuModule API, HIP defines a hipModule API that provides similar explicit control over code object management.

cuCtx API#

The driver API defines “Context” and “Devices” as separate entities. Contexts contain a single device, and a device can theoretically have multiple contexts. Each context contains a set of streams and events specific to the context. Historically, contexts also defined a unique address space for the GPU. This might no longer be the case in unified memory platforms, because the CPU and all the devices in the same process share a single unified address space. The Context APIs also provide a mechanism to switch between devices, which enables a single CPU thread to send commands to different GPUs. HIP and recent versions of the CUDA Runtime provide other mechanisms to accomplish this feat, for example, using streams or cudaSetDevice.

The CUDA runtime API unifies the Context API with the Device API. This simplifies the APIs and has little loss of functionality. This is because each context can contain a single device, and the benefits of multiple contexts have been replaced with other interfaces. HIP provides a Context API to facilitate easy porting from existing Driver code. In HIP, the Ctx functions largely provide an alternate syntax for changing the active device.

Most new applications preferentially use hipSetDevice or the stream APIs. Therefore, HIP has marked the hipCtx APIs as deprecated. Support for these APIs might not be available in future releases. For more details on deprecated APIs, see HIP deprecated runtime API functions.

HIP module and Ctx APIs#

Rather than present two separate APIs, HIP extends the HIP API with new APIs for modules and Ctx control.

hipModule API#

Like the CUDA driver API, the Module API provides additional control over how code is loaded, including options to load code from files or from in-memory pointers. NVCC and HIP-Clang target different architectures and use different code object formats. NVCC supports cubin or ptx files, while the HIP-Clang path uses the hsaco format. The external compilers which generate these code objects are responsible for generating and loading the correct code object for each platform. Notably, there is no fat binary format that can contain code for both NVCC and HIP-Clang platforms. The following table summarizes the formats used on each platform:

Module formats#

Format

APIs

NVCC

HIP-CLANG

Code object

hipModuleLoad, hipModuleLoadData

.cubin or PTX text

.hsaco

Fat binary

hipModuleLoadFatBin

.fatbin

.hip_fatbin

hipcc uses HIP-Clang or NVCC to compile host code. Both of these compilers can embed code objects into the final executable. These code objects are automatically loaded when the application starts. The hipModule API can be used to load additional code objects. When used this way, it extends the capability of the automatically loaded code objects. HIP-Clang enables both of these capabilities to be used together. Of course, it is possible to create a program with no kernels and no automatic loading.

For module API reference, visit Module management.

hipCtx API#

HIP provides a Ctx API as a thin layer over the existing device functions. The Ctx API can be used to set the current context or to query properties of the device associated with the context. The current context is implicitly used by other APIs, such as hipStreamCreate.

For context reference, visit Context management [deprecated].

HIPIFY translation of CUDA driver API#

The HIPIFY tools convert CUDA driver APIs for streams, events, modules, devices, memory management, context, and the profiler to the equivalent HIP calls. For example, cuEventCreate is translated to hipEventCreate. HIPIFY tools also convert error codes from the driver namespace and coding conventions to the equivalent HIP error code. HIP unifies the APIs for these common functions.

The memory copy API requires additional explanation. The CUDA driver includes the memory direction in the name of the API (cuMemcpyH2D), while the CUDA driver API provides a single memory copy API with a parameter that specifies the direction. It also supports a “default” direction where the runtime determines the direction automatically. HIP provides APIs with both styles, for example, hipMemcpyH2D as well as hipMemcpy. The first version might be faster in some cases because it avoids any host overhead to detect the different memory directions.

HIP defines a single error space and uses camel case for all errors (i.e. hipErrorInvalidValue).

For further information, visit the HIPIFY documentation.

Address spaces#

HIP-Clang defines a process-wide address space where the CPU and all devices allocate addresses from a single unified pool. This means addresses can be shared between contexts. Unlike the original CUDA implementation, a new context does not create a new address space for the device.

Using hipModuleLaunchKernel#

Both CUDA driver and runtime APIs define a function for launching kernels, called cuLaunchKernel or cudaLaunchKernel. The equivalent API in HIP is hipModuleLaunchKernel. The kernel arguments and the execution configuration (grid dimensions, group dimensions, dynamic shared memory, and stream) are passed as arguments to the launch function. The runtime API additionally provides the <<< >>> syntax for launching kernels, which resembles a special function call and is easier to use than the explicit launch API, especially when handling kernel arguments. However, this syntax is not standard C++ and is available only when NVCC is used to compile the host code.

Additional information#

HIP-Clang creates a primary context when the HIP API is called. So, in pure driver API code, HIP-Clang creates a primary context while HIP/NVCC has an empty context stack. HIP-Clang pushes the primary context to the context stack when it is empty. This can lead to subtle differences in applications which mix the runtime and driver APIs.

HIP-Clang implementation notes#

.hip_fatbin#

HIP-Clang links device code from different translation units together. For each device target, it generates a code object. clang-offload-bundler bundles code objects for different device targets into one fat binary, which is embedded as the global symbol __hip_fatbin in the .hip_fatbin section of the ELF file of the executable or shared object.

Initialization and termination functions#

HIP-Clang generates initialization and termination functions for each translation unit for host code compilation. The initialization functions call __hipRegisterFatBinary to register the fat binary embedded in the ELF file. They also call __hipRegisterFunction and __hipRegisterVar to register kernel functions and device-side global variables. The termination functions call __hipUnregisterFatBinary. HIP-Clang emits a global variable __hip_gpubin_handle of type void** with linkonce linkage and an initial value of 0 for each host translation unit. Each initialization function checks __hip_gpubin_handle and registers the fat binary only if __hip_gpubin_handle is 0. It saves the return value of __hip_gpubin_handle to __hip_gpubin_handle. This ensures that the fat binary is registered once. A similar check is performed in the termination functions.

Kernel launching#

HIP-Clang supports kernel launching using either the CUDA <<<>>> syntax, hipLaunchKernel, or hipLaunchKernelGGL. The last option is a macro which expands to the CUDA <<<>>> syntax by default. It can also be turned into a template by defining HIP_TEMPLATE_KERNEL_LAUNCH.

When the executable or shared library is loaded by the dynamic linker, the initialization functions are called. In the initialization functions, the code objects containing all kernels are loaded when __hipRegisterFatBinary is called. When __hipRegisterFunction is called, the stub functions are associated with the corresponding kernels in the code objects.

HIP-Clang implements two sets of APIs for launching kernels. By default, when HIP-Clang encounters the <<<>>> statement in the host code, it first calls hipConfigureCall to set up the threads and grids. It then calls the stub function with the given arguments. The stub function calls hipSetupArgument for each kernel argument, then calls hipLaunchByPtr with a function pointer to the stub function. In hipLaunchByPtr, the actual kernel associated with the stub function is launched.

NVCC implementation notes#

Interoperation between HIP and CUDA driver#

CUDA applications might want to mix CUDA driver code with HIP code (see the example below). This table shows the equivalence between CUDA and HIP types required to implement this interaction.

Equivalence table between HIP and CUDA types#

HIP type

CU Driver type

CUDA Runtime type

hipModule_t

CUmodule

hipFunction_t

CUfunction

hipCtx_t

CUcontext

hipDevice_t

CUdevice

hipStream_t

CUstream

cudaStream_t

hipEvent_t

CUevent

cudaEvent_t

hipArray

CUarray

cudaArray

Compilation options#

The hipModule_t interface does not support the cuModuleLoadDataEx function, which is used to control PTX compilation options. HIP-Clang does not use PTX, so it does not support these compilation options. In fact, HIP-Clang code objects contain fully compiled code for a device-specific instruction set and don’t require additional compilation as a part of the load step. The corresponding HIP function hipModuleLoadDataEx behaves like hipModuleLoadData on the HIP-Clang path (where compilation options are not used) and like cuModuleLoadDataEx on the NVCC path.

For example:

hipModule_t module;
void *imagePtr = ...; // Somehow populate data pointer with code object

const int numOptions = 1;
hipJitOption options[numOptions];
void *optionValues[numOptions];

options[0] = hipJitOptionMaxRegisters;
unsigned maxRegs = 15;
optionValues[0] = (void *)(&maxRegs);

// hipModuleLoadData(module, imagePtr) will be called on HIP-Clang path, JIT
// options will not be used, and cupModuleLoadDataEx(module, imagePtr,
// numOptions, options, optionValues) will be called on NVCC path
hipModuleLoadDataEx(module, imagePtr, numOptions, options, optionValues);

hipFunction_t k;
hipModuleGetFunction(&k, module, "myKernel");

CUmodule module;
void *imagePtr = ...; // Somehow populate data pointer with code object

const int numOptions = 1;
CUJit_option options[numOptions];
void *optionValues[numOptions];

options[0] = CU_JIT_MAX_REGISTERS;
unsigned maxRegs = 15;
optionValues[0] = (void *)(&maxRegs);

cuModuleLoadDataEx(module, imagePtr, numOptions, options, optionValues);

CUfunction k;
cuModuleGetFunction(&k, module, "myKernel");

The sample below shows how to use hipModuleGetFunction.

#include <hip/hip_runtime.h>
#include <hip/hip_runtime_api.h>

#include <vector>

int main() {

    size_t elements = 64*1024;
    size_t size_bytes = elements * sizeof(float);

    std::vector<float> A(elements), B(elements);

    // On NVIDIA platforms the driver runtime needs to be initiated
    #ifdef __HIP_PLATFORM_NVIDIA__
    hipInit(0);
    hipDevice_t device;
    hipCtx_t context;
    HIPCHECK(hipDeviceGet(&device, 0));
    HIPCHECK(hipCtxCreate(&context, 0, device));
    #endif

    // Allocate device memory
    hipDeviceptr_t d_A, d_B;
    HIPCHECK(hipMalloc(&d_A, size_bytes));
    HIPCHECK(hipMalloc(&d_B, size_bytes));

    // Copy data to device
    HIPCHECK(hipMemcpyHtoD(d_A, A.data(), size_bytes));
    HIPCHECK(hipMemcpyHtoD(d_B, B.data(), size_bytes));

    // Load module
    hipModule_t Module;
    // For AMD the module file has to contain architecture specific object codee
    // For NVIDIA the module file has to contain PTX, found in e.g. "vcpy_isa.ptx"
    HIPCHECK(hipModuleLoad(&Module, "vcpy_isa.co"));
    // Get kernel function from the module via its name
    hipFunction_t Function;
    HIPCHECK(hipModuleGetFunction(&Function, Module, "hello_world"));

    // Create buffer for kernel arguments
    std::vector<void*> argBuffer{&d_A, &d_B};
    size_t arg_size_bytes = argBuffer.size() * sizeof(void*);

    // Create configuration passed to the kernel as arguments
    void* config[] = {HIP_LAUNCH_PARAM_BUFFER_POINTER, argBuffer.data(),
                      HIP_LAUNCH_PARAM_BUFFER_SIZE, &arg_size_bytes, HIP_LAUNCH_PARAM_END};

    int threads_per_block = 128;
    int blocks = (elements + threads_per_block - 1) / threads_per_block;

    // Actually launch kernel
    HIPCHECK(hipModuleLaunchKernel(Function, blocks, 1, 1, threads_per_block, 1, 1, 0, 0, NULL, config));

    HIPCHECK(hipMemcpyDtoH(A.data(), d_A, elements));
    HIPCHECK(hipMemcpyDtoH(B.data(), d_B, elements));

    #ifdef __HIP_PLATFORM_NVIDIA__
    HIPCHECK(hipCtxDetach(context));
    #endif

    HIPCHECK(hipFree(d_A));
    HIPCHECK(hipFree(d_B));

    return 0;
}

HIP module and texture Driver API#

HIP supports texture driver APIs. However, texture references must be declared within the host scope. The following code demonstrates the use of texture references for the __HIP_PLATFORM_AMD__ platform.

// Code to generate code object

#include "hip/hip_runtime.h"
extern texture<float, 2, hipReadModeElementType> tex;

__global__ void tex2dKernel(hipLaunchParm lp, float *outputData, int width,
                            int height) {
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    int y = blockIdx.y * blockDim.y + threadIdx.y;
    outputData[y * width + x] = tex2D(tex, x, y);
}

// Host code:

texture<float, 2, hipReadModeElementType> tex;

  void myFunc ()
  {
      // ...

      textureReference* texref;
      hipModuleGetTexRef(&texref, Module1, "tex");
      hipTexRefSetAddressMode(texref, 0, hipAddressModeWrap);
      hipTexRefSetAddressMode(texref, 1, hipAddressModeWrap);
      hipTexRefSetFilterMode(texref, hipFilterModePoint);
      hipTexRefSetFlags(texref, 0);
      hipTexRefSetFormat(texref, HIP_AD_FORMAT_FLOAT, 1);
      hipTexRefSetArray(texref, array, HIP_TRSA_OVERRIDE_FORMAT);

    // ...
  }

Driver entry point access#

Starting from HIP version 6.2.0, support for Driver Entry Point Access is available when using CUDA 12.0 or newer. This feature allows developers to directly interact with the CUDA driver API, providing more control over GPU operations.

Driver Entry Point Access provides several features:

  • Retrieving the address of a runtime function

  • Requesting the default stream version on a per-thread basis

  • Accessing new HIP features on older toolkits with a newer driver

For driver entry point access reference, visit hipGetProcAddress().

Address retrieval#

The hipGetProcAddress() function can be used to obtain the address of a runtime function. This is demonstrated in the following example:

#include <hip/hip_runtime.h>
#include <hip/hip_runtime_api.h>

#include <iostream>

typedef hipError_t (*hipInit_t)(unsigned int);

int main() {
    // Initialize the HIP runtime
    hipError_t res = hipInit(0);
    if (res != hipSuccess) {
        std::cerr << "Failed to initialize HIP runtime." << std::endl;
        return 1;
    }

    // Get the address of the hipInit function
    hipInit_t hipInitFunc;
    int hipVersion = HIP_VERSION; // Use the HIP version defined in hip_runtime_api.h
    uint64_t flags = 0; // No special flags
    hipDriverProcAddressQueryResult symbolStatus;

    res = hipGetProcAddress("hipInit", (void**)&hipInitFunc, hipVersion, flags, &symbolStatus);
    if (res != hipSuccess) {
        std::cerr << "Failed to get address of hipInit()." << std::endl;
        return 1;
    }

    // Call the hipInit function using the obtained address
    res = hipInitFunc(0);
    if (res == hipSuccess) {
        std::cout << "HIP runtime initialized successfully using hipGetProcAddress()." << std::endl;
    } else {
        std::cerr << "Failed to initialize HIP runtime using hipGetProcAddress()." << std::endl;
    }

    return 0;
}

Per-thread default stream version request#

HIP offers functionality similar to CUDA for managing streams on a per-thread basis. By using hipStreamPerThread, each thread can independently manage its default stream, simplifying operations. The following example demonstrates how this feature enhances performance by reducing contention and improving efficiency.

#include <hip/hip_runtime.h>

#include <iostream>

int main() {
    // Initialize the HIP runtime
    hipError_t res = hipInit(0);
    if (res != hipSuccess) {
        std::cerr << "Failed to initialize HIP runtime." << std::endl;
        return 1;
    }

    // Get the per-thread default stream
    hipStream_t stream = hipStreamPerThread;

    // Use the stream for some operation
    // For example, allocate memory on the device
    void* d_ptr;
    size_t size = 1024;
    res = hipMalloc(&d_ptr, size);
    if (res != hipSuccess) {
        std::cerr << "Failed to allocate memory." << std::endl;
        return 1;
    }

    // Perform some operation using the stream
    // For example, set memory on the device
    res = hipMemsetAsync(d_ptr, 0, size, stream);
    if (res != hipSuccess) {
        std::cerr << "Failed to set memory." << std::endl;
        return 1;
    }

    // Synchronize the stream
    res = hipStreamSynchronize(stream);
    if (res != hipSuccess) {
        std::cerr << "Failed to synchronize stream." << std::endl;
        return 1;
    }

    std::cout << "Operation completed successfully using per-thread default stream." << std::endl;

    // Free the allocated memory
    hipFree(d_ptr);

    return 0;
}

Accessing new HIP features with a newer driver#

HIP is designed to be forward compatible, allowing newer features to be utilized with older toolkits, provided a compatible driver is present. Feature support can be verified through runtime API functions and version checks. This approach ensures that applications can benefit from new features and improvements in the HIP runtime without needing to be recompiled with a newer toolkit. The function hipGetProcAddress() enables dynamic querying and the use of newer functions offered by the HIP runtime, even if the application was built with an older toolkit.

An example is provided for a hypothetical foo() function.

// Get the address of the foo function
foo_t fooFunc;
int hipVersion = 60300000; // Use an own HIP version number (e.g. 6.3.0)
uint64_t flags = 0; // No special flags
hipDriverProcAddressQueryResult symbolStatus;

res = hipGetProcAddress("foo", (void**)&fooFunc, hipVersion, flags, &symbolStatus);

The HIP version number is defined as an integer:

HIP_VERSION=HIP_VERSION_MAJOR * 10000000 + HIP_VERSION_MINOR * 100000 + HIP_VERSION_PATCH
Contents