Unified memory#

In conventional architectures, CPUs and GPUs have dedicated memory like Random Access Memory (RAM) and Video Random Access Memory (VRAM). This architectural design, while effective, can be limiting in terms of memory capacity and bandwidth, as continuous memory copying is required to allow the processors to access the appropriate data. New architectural features like Heterogeneous System Architectures (HSA) and Unified Memory (UM) help avoid these limitations and promise increased efficiency and innovation.

Unified memory#

Unified Memory is a single memory address space accessible from any processor within a system. This setup simplifies memory management processes and enables applications to allocate data that can be read or written by code running on either CPUs or GPUs. The Unified memory model is shown in the following figure.

../_images/um.svg

AMD Accelerated Processing Unit (APU) is a typical example of a Unified Memory Architecture. On a single die, a central processing unit (CPU) is combined with an integrated graphics processing unit (iGPU), and both have access to a high-bandwidth memory (HBM) module named Unified Memory. The CPU enables high-performance, low-latency operations, while the GPU is optimized for high throughput (data processed by unit time).

System requirements#

Unified memory is supported on Linux by all modern AMD GPUs from the Vega series onward. Unified memory management can be achieved with managed memory allocation and, for the latest GPUs, with a system allocator.

The table below lists the supported allocators. The allocators are described in the next section.

Supported Unified Memory Allocators#

Architecture

hipMallocManaged()

__managed__

malloc()

MI200, MI300 Series

1

MI100

RDNA (Navi) Series

GCN5 (Vega) Series

✅: Supported

❌: Unsupported

1 Works only with XNACK=1. First GPU access causes recoverable page-fault. For more details, visit GPU memory.

Unified memory programming models#

Showcasing various unified memory programming models, the model availability depends on your architecture. For more information, see System requirements and Checking unified memory management support.

  • HIP managed memory allocation API:

    The hipMallocManaged() is a dynamic memory allocator available on all GPUs with unified memory support. For more details, visit HIP managed memory allocation API.

  • HIP managed variables:

    The __managed__ declaration specifier, which serves as its counterpart, is supported on all modern AMD cards and can be utilized for static allocation.

  • System allocation API:

    Starting with the AMD MI300 series, the malloc() system allocator allows you to reserve unified memory. The system allocator is more versatile and offers an easy transition from a CPU written C++ code to a HIP code as the same system allocation API is used.

Checking unified memory management support#

Some device attributes can offer information about which Unified memory programming models are supported. The attribute value is 1 if the functionality is supported, and 0 if it is not supported.

Device attributes for unified memory management#

attribute

description

hipDeviceAttributeManagedMemory

unified addressing is supported

hipDeviceAttributeConcurrentManagedAccess

full managed memory support, concurrent access is supported

hipDeviceAttributePageableMemoryAccess

both managed and system memory allocation API is supported

The following examples show how to use device attributes:

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

int main() {
    int d;
    hipGetDevice(&d);

    int is_cma = 0;
    hipDeviceGetAttribute(&is_cma, hipDeviceAttributeConcurrentManagedAccess, d);
    std::cout << "HIP Managed Memory: "
              << (is_cma == 1 ? "is" : "NOT")
              << " supported" << std::endl;
    return 0;
}

Example for unified memory management#

The following example shows how to use unified memory management with hipMallocManaged(), function, with __managed__ attribute for static allocation and standard malloc() allocation. For comparison, the Explicit Memory Management example is presented in the last tab.

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

// Addition of two values.
__global__ void add(int *a, int *b, int *c) {
    *c = *a + *b;
}

int main() {
    int *a, *b, *c;

    // Allocate memory for a, b and c that is accessible to both device and host codes.
    hipMallocManaged(&a, sizeof(*a));
    hipMallocManaged(&b, sizeof(*b));
    hipMallocManaged(&c, sizeof(*c));

    // Setup input values.
    *a = 1;
    *b = 2;

    // Launch add() kernel on GPU.
    hipLaunchKernelGGL(add, dim3(1), dim3(1), 0, 0, a, b, c);

    // Wait for GPU to finish before accessing on host.
    hipDeviceSynchronize();

    // Prints the result.
    std::cout << *a << " + " << *b << " = " << *c << std::endl;

    // Cleanup allocated memory.
    hipFree(a);
    hipFree(b);
    hipFree(c);

    return 0;
}
#include <hip/hip_runtime.h>
#include <iostream>

// Addition of two values.
__global__ void add(int *a, int *b, int *c) {
    *c = *a + *b;
}

// Declare a, b and c as static variables.
__managed__ int a, b, c;

int main() {
    // Setup input values.
    a = 1;
    b = 2;

    // Launch add() kernel on GPU.
    hipLaunchKernelGGL(add, dim3(1), dim3(1), 0, 0, &a, &b, &c);

    // Wait for GPU to finish before accessing on host.
    hipDeviceSynchronize();

    // Prints the result.
    std::cout << a << " + " << b << " = " << c << std::endl;

    return 0;
}
#include <hip/hip_runtime.h>
#include <iostream>

// Addition of two values.
__global__ void add(int* a, int* b, int* c) {
    *c = *a + *b;
}

int main() {
    int* a, * b, * c;

    // Allocate memory for a, b, and c.
    a = (int*)malloc(sizeof(*a));
    b = (int*)malloc(sizeof(*b));
    c = (int*)malloc(sizeof(*c));

    // Setup input values.
    *a = 1;
    *b = 2;

    // Launch add() kernel on GPU.
    hipLaunchKernelGGL(add, dim3(1), dim3(1), 0, 0, a, b, c);

    // Wait for GPU to finish before accessing on host.
    hipDeviceSynchronize();

    // Prints the result.
    std::cout << *a << " + " << *b << " = " << *c << std::endl;

    // Cleanup allocated memory.
    free(a);
    free(b);
    free(c);

    return 0;
}
#include <hip/hip_runtime.h>
#include <iostream>

// Addition of two values.
__global__ void add(int *a, int *b, int *c) {
    *c = *a + *b;
}

int main() {
    int a, b, c;
    int *d_a, *d_b, *d_c;

    // Setup input values.
    a = 1;
    b = 2;

    // Allocate device copies of a, b and c.
    hipMalloc(&d_a, sizeof(*d_a));
    hipMalloc(&d_b, sizeof(*d_b));
    hipMalloc(&d_c, sizeof(*d_c));

    // Copy input values to device.
    hipMemcpy(d_a, &a, sizeof(*d_a), hipMemcpyHostToDevice);
    hipMemcpy(d_b, &b, sizeof(*d_b), hipMemcpyHostToDevice);

    // Launch add() kernel on GPU.
    hipLaunchKernelGGL(add, dim3(1), dim3(1), 0, 0, d_a, d_b, d_c);

    // Copy the result back to the host.
    hipMemcpy(&c, d_c, sizeof(*d_c), hipMemcpyDeviceToHost);

    // Cleanup allocated memory.
    hipFree(d_a);
    hipFree(d_b);
    hipFree(d_c);

    // Prints the result.
    std::cout << a << " + " << b << " = " << c << std::endl;

    return 0;
}

Using unified memory management (UMM)#

Unified memory management (UMM) is a feature that can simplify the complexities of memory management in GPU computing. It is particularly useful in heterogeneous computing environments with heavy memory usage with both a CPU and a GPU, which would require large memory transfers. Here are some areas where UMM can be beneficial:

  • Simplification of Memory Management:

    UMM can help to simplify the complexities of memory management. This can make it easier for developers to write code without worrying about memory allocation and deallocation details.

  • Data Migration:

    UMM allows for efficient data migration between the host (CPU) and the device (GPU). This can be particularly useful for applications that need to move data back and forth between the device and host.

  • Improved Programming Productivity:

    As a positive side effect, UMM can reduce the lines of code, thereby improving programming productivity.

In HIP, pinned memory allocations are coherent by default. Pinned memory is host memory mapped into the address space of all GPUs, meaning that the pointer can be used on both host and device. Using pinned memory instead of pageable memory on the host can improve bandwidth.

While UMM can provide numerous benefits, it’s important to be aware of the potential performance overhead associated with UMM. You must thoroughly test and profile your code to ensure it’s the most suitable choice for your use case.

Unified memory HIP runtime hints for the better performance#

Unified memory HIP runtime hints can help improve the performance of your code if you know your code’s ability and infrastructure. Some hint techniques are presented in this section.

The hint functions can set actions on a selected device, which can be identified by hipGetDeviceProperties(&prop, device_id). There are two special device_id values:

  • hipCpuDeviceId = -1 means that the advised device is the CPU.

  • hipInvalidDeviceId = -2 means that the device is invalid.

For the best performance, profile your application to optimize the utilization of HIP runtime hints.

Data prefetching#

Data prefetching is a technique used to improve the performance of your application by moving data closer to the processing unit before it’s actually needed.

// Addition of two values.
__global__ void add(int *a, int *b, int *c) {
    *c = *a + *b;
}

int main() {
    int *a, *b, *c;
    int deviceId;
    hipGetDevice(&deviceId); // Get the current device ID

    // Allocate memory for a, b and c that is accessible to both device and host codes.
    hipMallocManaged(&a, sizeof(*a));
    hipMallocManaged(&b, sizeof(*b));
    hipMallocManaged(&c, sizeof(*c));

    // Setup input values.
    *a = 1;
    *b = 2;

    // Prefetch the data to the GPU device.
    hipMemPrefetchAsync(a, sizeof(*a), deviceId, 0);
    hipMemPrefetchAsync(b, sizeof(*b), deviceId, 0);
    hipMemPrefetchAsync(c, sizeof(*c), deviceId, 0);

    // Launch add() kernel on GPU.
    hipLaunchKernelGGL(add, dim3(1), dim3(1), 0, 0, a, b, c);

    // Wait for GPU to finish before accessing on host.
    hipDeviceSynchronize();

    // Prefetch the result back to the CPU.
    hipMemPrefetchAsync(c, sizeof(*c), hipCpuDeviceId, 0);

    // Wait for the prefetch operations to complete.
    hipDeviceSynchronize();

    // Prints the result.
    std::cout << *a << " + " << *b << " = " << *c << std::endl;

    // Cleanup allocated memory.
    hipFree(a);
    hipFree(b);
    hipFree(c);

    return 0;
}

Remember to check the return status of hipMemPrefetchAsync() to ensure that the prefetch operations are completed successfully.

Memory advice#

The effectiveness of hipMemAdvise() comes from its ability to inform the runtime system of the developer’s intentions regarding memory usage. When the runtime system has knowledge of the expected memory access patterns, it can make better decisions about data placement and caching, leading to more efficient execution of the application. However, the actual impact on performance can vary based on the specific use case and the hardware architecture.

For the description of hipMemAdvise() and the detailed list of advice, visit the HIP managed memory allocation API.

Here is the updated version of the example above with memory advice.

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

// Addition of two values.
__global__ void add(int *a, int *b, int *c) {
    *c = *a + *b;
}

int main() {
    int *a, *b, *c;

    // Allocate memory for a, b, and c accessible to both device and host codes.
    hipMallocManaged(&a, sizeof(*a));
    hipMallocManaged(&b, sizeof(*b));
    hipMallocManaged(&c, sizeof(*c));

    // Set memory advice for a, b, and c to be accessed by the CPU.
    hipMemAdvise(a, sizeof(*a), hipMemAdviseSetPreferredLocation, hipCpuDeviceId);
    hipMemAdvise(b, sizeof(*b), hipMemAdviseSetPreferredLocation, hipCpuDeviceId);
    hipMemAdvise(c, sizeof(*c), hipMemAdviseSetPreferredLocation, hipCpuDeviceId);

    // Additionally, set memory advice for a, b, and c to be read mostly from the device 0.
    constexpr int device = 0;
    hipMemAdvise(a, sizeof(*a), hipMemAdviseSetReadMostly, device);
    hipMemAdvise(b, sizeof(*b), hipMemAdviseSetReadMostly, device);
    hipMemAdvise(c, sizeof(*c), hipMemAdviseSetReadMostly, device);

    // Setup input values.
    *a = 1;
    *b = 2;

    // Launch add() kernel on GPU.
    hipLaunchKernelGGL(add, dim3(1), dim3(1), 0, 0, a, b, c);

    // Wait for GPU to finish before accessing on host.
    hipDeviceSynchronize();

    // Prints the result.
    std::cout << *a << " + " << *b << " = " << *c << std::endl;

    // Cleanup allocated memory.
    hipFree(a);
    hipFree(b);
    hipFree(c);

    return 0;
}

Memory range attributes#

Memory Range attributes allow you to query attributes of a given memory range.

The hipMemRangeGetAttribute() is added to the example to query the hipMemRangeAttributeReadMostly attribute of the memory range pointed to by a. The result is stored in attributeValue and then printed out.

For more details, visit the HIP managed memory allocation API.

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

// Addition of two values.
__global__ void add(int *a, int *b, int *c) {
    *c = *a + *b;
}

int main() {
    int *a, *b, *c;
    unsigned int attributeValue;
    constexpr size_t attributeSize = sizeof(attributeValue);

    // Allocate memory for a, b and c that is accessible to both device and host codes.
    hipMallocManaged(&a, sizeof(*a));
    hipMallocManaged(&b, sizeof(*b));
    hipMallocManaged(&c, sizeof(*c));

    // Setup input values.
    *a = 1;
    *b = 2;

    // Launch add() kernel on GPU.
    hipLaunchKernelGGL(add, dim3(1), dim3(1), 0, 0, a, b, c);

    // Wait for GPU to finish before accessing on host.
    hipDeviceSynchronize();

    // Query an attribute of the memory range.
    hipMemRangeGetAttribute(&attributeValue,
                            attributeSize,
                            hipMemRangeAttributeReadMostly,
                            a,
                            sizeof(*a));

    // Prints the result.
    std::cout << *a << " + " << *b << " = " << *c << std::endl;
    std::cout << "The queried attribute value is: " << attributeValue << std::endl;

    // Cleanup allocated memory.
    hipFree(a);
    hipFree(b);
    hipFree(c);

    return 0;
}

Asynchronously attach memory to a stream#

The hipStreamAttachMemAsync function would be able to asynchronously attach memory to a stream, which can help concurrent execution when using streams.

Currently, this function is a no-operation (NOP) function on AMD GPUs. It simply returns success after the runtime memory validation passed. This function is necessary on Microsoft Windows, and UMM is not supported on this operating system with AMD GPUs at the moment.