.. meta:: :description: This chapter explains the HIP programming model, the contract between the programmer and the compiler/runtime executing the code, how it maps to the hardware. :keywords: ROCm, HIP, CUDA, API design, programming model .. _programming_model: ******************************************************************************* Introduction to the HIP programming model ******************************************************************************* The HIP programming model enables mapping data-parallel C/C++ algorithms to massively parallel SIMD (Single Instruction, Multiple Data) architectures like GPUs. HIP supports many imperative languages, such as Python via PyHIP, but this document focuses on the original C/C++ API of HIP. While GPUs may be capable of running applications written for CPUs if properly ported and compiled, it would not be an efficient use of GPU resources. GPUs fundamentally differ from CPUs and should be used accordingly to achieve optimum performance. A basic understanding of the underlying device architecture helps you make efficient use of HIP and general purpose graphics processing unit (GPGPU) programming in general. The following topics introduce you to the key concepts of GPU-based programming and the HIP programming model. Hardware differences: CPU vs GPU ================================ CPUs and GPUs have been designed for different purposes. CPUs quickly execute a single thread, decreasing the time for a single operation while increasing the number of sequential instructions that can be executed. This includes fetching data and reducing pipeline stalls where the ALU has to wait for previous instructions to finish. .. figure:: ../data/understand/programming_model/cpu-gpu-comparison.svg :alt: Diagram depicting the differences between CPU and GPU hardware. The CPU block shows four large processing cores, lists Large Cache per Core, and High Clock Speed of 3 to 5 gigahertz. The GPU block shows 42 smaller processing cores, lists Shared Memory across Cores, and Lower Clock Speeds of 1 to 2 gigahertz. Differences in CPUs and GPUs With CPUs, the goal is to quickly process operations. CPUs provide low-latency processing for serial instructions. On the other hand, GPUs have been designed to execute many similar commands, or threads, in parallel, achieving higher throughput. Latency is the time between starting an operation and receiving its result, such as 2 ns, while throughput is the rate of completed operations, for example, operations per second. For the GPU, the objective is to process as many operations in parallel, rather than to finish a single instruction quickly. GPUs in general are made up of basic building blocks called :ref:`compute units (CUs) `, that execute the threads of a kernel. As described in :ref:`hardware_implementation`, these CUs provide the necessary resources for the threads: the Arithmetic Logical Units (ALUs), register files, caches and shared memory for efficient communication between the threads. The following describes a few hardware differences between CPUs and GPUs: * CPU: - Optimized for sequential processing with a few powerful cores (4-64 typically) - High clock speeds (3-5 GHz) - One register file per thread. On modern CPUs you have at most 2 register files per core, called hyperthreading. - One ALU executing the thread. - Designed to quickly execute instructions of the same thread. - Complex branch prediction. - Large L1/L2 cache per core, shared by fewer threads (maximum of 2 when hyperthreading is available). - A disadvantage is switching execution from one thread to another (or context switching) takes a considerable amount of time: the ALU pipeline needs to be emptied, the register file has to be written to memory to free the register for another thread. * GPU: - Designed for parallel processing with many simpler cores (hundreds/thousands) - Lower clock speeds (1-2 GHz) - Streamlined control logic - Small caches, more registers - Register files are shared among threads. The number of threads that can be run in parallel depends on the registers needed per thread. - Multiple ALUs execute a collection of threads having the same operations, also known as a :ref:`warp ` or wavefront. This is called single-instruction, multiple threads (SIMT) operation as described in :ref:`programming_model_simt`. - The collection of ALUs is called SIMD. SIMDs are an extension to the hardware architecture that allows a `single instruction` to concurrently operate on `multiple data` inputs. - For branching threads where conditional instructions lead to thread divergence, ALUs still process the full warp, but the result for divergent threads is masked out. This leads to wasted ALU cycles and should be a consideration in your programming. Keep instructions consistent and leave conditionals out of threads. - The advantage for GPUs is that context switching is easy. All threads that run on a core/compute unit have their registers on the compute unit, so they don't need to be stored to global memory, and each cycle one instruction from any wavefront that resides on the compute unit can be issued. When programming for a heterogeneous system, which incorporates CPUs and GPUs, you must write your program to take advantage of the strengths of the available hardware. Use the CPU for tasks that require complex logic with conditional branching, to reduce the time to reach a decision. Use the GPU for parallel operations of the same instruction across large datasets, with little branching, where the volume of operations is the key. .. _heterogeneous_programming: Heterogeneous programming ========================= The HIP programming model has two execution contexts. The main application starts on the CPU, or the *host* processor, and compute kernels are launched on the *device* such as `Instinct accelerators `_ or AMD GPUs. The host execution is defined by the C++ abstract machine, while device execution follows the :ref:`SIMT model` of HIP. These two execution contexts are signified by the ``__host__`` and ``__global__`` (or ``__device__``) decorators in HIP program code. There are a few key differences between the two contexts: * The C++ abstract machine assumes a unified memory address space, meaning that one can always access any given address in memory (assuming the absence of data races). HIP however introduces several memory namespaces, an address from one means nothing in another. Moreover, not all address spaces are accessible from all contexts. Every CU has an instance of storage backing the namespace ``__shared__``. Even if the host were to have access to these regions of memory, the performance benefits of the segmented memory subsystem are supported by the inability of asynchronous access from the host. * Not all C++ language features map cleanly to typical GPU device architectures. Some C++ features have poor latency when implemented on GPU devices, therefore they are forbidden in device contexts to avoid using features that unexpectedly decimate the program's performance. Offload devices targeted by HIP aren't general purpose devices, at least not in the sense that a CPU is. HIP focuses on data parallel computations and as such caters to throughput optimized architectures, such as GPUs or accelerators derived from GPU architectures. * Asynchronicity is at the forefront of the HIP API. Computations launched on the device execute asynchronously with respect to the host, and it is the user's responsibility to synchronize their data dispatch/fetch with computations on the device. .. note:: HIP performs implicit synchronization on occasions, unlike some APIs where the responsibility for synchronization is left to the user. Host programming ---------------- In heterogeneous programming, the CPU is available for processing operations but the host application has the additional task of managing data and computation exchanges between the CPU (host) and GPU (device). The host acts as the application manager, coordinating the overall workflow and directing operations to the appropriate context, handles data preparation and data transfers, and manages GPU tasks and synchronization. Here is a typical sequence of operations: 1. Initialize the HIP runtime and select the GPU: As described in :ref:`initialization`, refers to identifying and selecting a target GPU, setting up a context to let the CPU interact with the GPU. 2. Data preparation: As discussed in :ref:`memory_management`, this includes allocating the required memory on the host and device, preparing input data and transferring it from the host to the device. The data is both transferred to the device, and passed as an input parameter when launching the kernel. 3. Configure and launch the kernel on the GPU: As described in :ref:`device_program`, this defines kernel configurations and arguments, launches kernel to run on the GPU device using the triple chevron syntax or appropriate API call (for example ``hipLaunchKernelGGL``). On the GPU, multiple kernels can run on streams, with a queue of operations. Within the same stream, operations run in the order they were issued, but on multiple streams operations are independent and can execute concurrently. In the HIP runtime, kernels run on the default stream when one is not specified, but specifying a stream for the kernel lets you increase concurrency in task scheduling and resource utilization, and launch and manage multiple kernels from the host program. 4. Synchronization: As described in :ref:`asynchronous_how-to`, kernel execution occurs in the context of device streams, specifically the default (`0`) stream. You can use streams and events to manage task dependencies, overlap computation with data transfers, and manage asynchronous processes to ensure proper sequencing of operations. Wait for events or streams to finish execution and transfer results from the GPU back to the host. 5. Error handling: As described in :ref:`error_handling`, you should catch and handle potential errors from API calls, kernel launches, or memory operations. For example, use ``hipGetErrorString`` to retrieve error messages. 6. Cleanup and resource management: Validate results, clean up GPU contexts and resources, and free allocated memory on the host and devices. This structure allows for efficient use of GPU resources and facilitates the acceleration of compute-intensive tasks while keeping the host CPU available for other tasks. .. figure:: ../data/understand/programming_model/host-device-flow.svg :alt: Diagram depicting a host CPU and device GPU rectangles of varying color. There are arrows pointing between the rectangles showing from the Host to the Device the initialization, data transfer, and Kernel execution steps, and from the Device back to the Host the returning results. Interaction of Host and Device in a GPU application HIP Runtime API ~~~~~~~~~~~~~~~ The HIP Runtime API provides a high-level C-style interface that abstracts the lower-level ROCr Runtime, simplifying GPU programming on AMD hardware. It offers functions for memory management, kernel launches, synchronization, and device control, enabling developers to write GPU-accelerated applications in a portable, C++-friendly way. The Runtime API serves the same role within ROCm as the CUDA Runtime API does in NVIDIA's stack. It simplifies GPU programming by abstracting explicit device and context management handled by the HSA/ROCr driver layer. Developers interact with familiar, high-level constructs such as :cpp:func:`hipMalloc`, :cpp:func:`hipMemcpy`, and ``hipLaunchKernelGGL``, while the runtime handles resource allocation, queue management, and synchronization transparently. The Runtime API can be linked either statically or dynamically, with the shared object typically named ``libamdhip64.so`` on Linux systems. It is open source and maintained as part of the ROCm ecosystem, allowing inspection, extension, and integration into custom build environments. .. _device_program: Device programming ------------------ The device or kernel program acts as a worker on the GPU application, distributing operations to be handled quickly and efficiently. Launching a kernel in the host application starts the kernel program running on the GPU, defining the parallel operations that repeat the same instructions across many datasets. Understanding how the kernel works and the processes involved is essential to writing efficient GPU applications. Threads, blocks, and grids provide a hierarchical approach to parallel operations. Understanding the thread hierarchy is critical to distributing work across the available CUs, managing parallel operations, and optimizing memory access. The general flow of the kernel program looks like this: 1. Thread Grouping: As described in :ref:`inherent_thread_model`, threads are organized into a hierarchy consisting of threads, which are individual instances of parallel operations, blocks that group the threads, and grids that group blocks into the kernel. Each thread runs an instance of the kernel in parallel with other threads in the block. 2. Indexing: The kernel computes the unique index for each thread to access the relevant data to be processed by the thread. 3. Data Fetch: Threads fetch input data from memory previously transferred from the host to the device. As described in :ref:`memory_hierarchy`, the hierarchy of threads is influenced by the memory subsystem of GPUs. The memory hierarchy includes local memory per-thread with very fast access, shared memory for the block of threads which also supports quick access, and larger amounts of global memory visible to the whole kernel,but accesses are expensive due to high latency. Understanding the memory model is a key concept for kernel programming. 4. Computation: Threads perform the required computations on the input data, and generate any needed output. Each thread of the kernel runs the same instruction simultaneously on the different datasets. This sometimes require multiple iterations when the number of operations exceeds the resources of the CU. 5. Synchronization: When needed, threads synchronize within their block to ensure correct results when working with shared memory. Kernels are parallel programs that execute the same instruction set across multiple threads, organized in :ref:`warps `, as described below and as demonstrated in the `Hello World tutorial `_ or :doc:`../tutorial/saxpy`. However, heterogeneous GPU applications can also become quite complex, managing hundreds, thousands, or hundreds of thousands of operations with repeated data transfers between host and device to support massive parallelization, using multiple streams to manage concurrent asynchronous operations, using rich libraries of functions optimized for GPU hardware as described in the `ROCm documentation `_. .. _programming_model_simt: Single instruction multiple threads (SIMT) ========================================== The HIP kernel code, written as a series of scalar instructions for multiple threads with different thread indices, gets mapped to the SIMD units of the GPUs. Every single instruction, which is executed for every participating thread of a kernel, gets mapped to the SIMD. This is done by grouping threads into warps, which contain as many threads as there are physical lanes in a SIMD, and issuing that instruction to the SIMD for every warp of a kernel. Ideally, the SIMD is always fully utilized. However, if the number of threads can't be evenly divided by the warpSize, then the unused lanes are masked out from the corresponding SIMD execution. .. _simt: .. figure:: ../data/understand/programming_model/simt-execution.svg :alt: Diagram depicting the SIMT execution model. There is a red rectangle which contains the expression a[i] = b[i] + c[i], and below that four arrows that point to Thread 0,1,2, and 3. Each thread contains different values for b, c, and a, showing the parallel operations of this equation. Instruction flow of a sample SIMT program A kernel follows the same C++ rules as the functions on the host, but it has a special ``__global__`` label to mark it for execution on the device, as shown in the following example: .. code-block:: cpp __global__ void AddKernel(float* a, const float* b) { int global_idx = threadIdx.x + blockIdx.x * blockDim.x; a[global_idx] += b[global_idx]; } One of the first things you might notice is the usage of the special ``threadIdx``, ``blockIdx`` and ``blockDim`` variables. Unlike normal C++ host functions, a kernel is not launched once, but as often as specified by the user. Each of these instances is a separate thread, with its own values for ``threadIdx``, ``blockIdx`` and ``blockDim``. The kernel program is launched from the host application using a language extension called the triple chevron syntax, which looks like the following: .. code-block:: cpp AddKernel<<>>(a, b); Inside the angle brackets, provide the following: * The number of blocks to launch, which defines the grid size (relating to blockDim). * The number of threads in a block, which defines the block size (relating to blockIdx). * The amount of shared memory to allocate by the host, not specified above. * The device stream to enqueue the operation on, not specified above so the default stream is used. .. note:: The kernel can also be launched through other methods, such as the ``hipLaunchKernel()`` function. Here, the total number of threads launched for the ``AddKernel`` program is defined by ``number_of_blocks * threads_per_block``. You define these values when launching the kernel program to address the problem to be solved with the available resources within the system. In other words, the thread configuration is customized to the needs of the operations and the available hardware. For comparison, the ``AddKernel`` program could be written in plain C++ as a ``FOR`` loop: .. code-block:: cpp for(int i = 0; i < (number_of_blocks * threads_per_block); ++i){ a[i] += b[i]; } In HIP, lanes of the SIMD architecture are fed by mapping threads of a SIMT execution, one thread down each lane of an SIMD engine. Execution parallelism usually isn't exploited from the width of the built-in vector types, but across multiple threads via the thread ID constants ``threadIdx.x``, ``blockIdx.x``, etc. .. _inherent_thread_model: Hierarchical thread model ------------------------- The thread hierarchy defines how parallel work is structured and executed across AMD GPUs, from single threads up to the full device. All threads of a kernel are uniquely identified by a set of integral values called thread IDs. The hierarchy consists of three levels: threads, blocks, and grids. * Threads are single instances of kernel operations, running concurrently * Blocks group threads together and enable cooperation and shared memory * Grids define the number of thread blocks for a single kernel launch * Blocks and grids can be defined in three dimensions (``x``, ``y``, ``z``) * By default, the Y and Z dimensions are set to 1 This hierarchy maps directly onto AMD hardware: * Threads execute on SIMD lanes * :ref:`Work-groups ` occupy :ref:`compute units ` * :ref:`Grids ` utilize all available CUs across the GPU The combined values represent the thread index, and relate to the sequence that the threads execute. The thread hierarchy is integral to how AMD GPUs operate, and is depicted in the following figure. .. figure:: ../data/understand/programming_model/thread_hierarchy.svg :alt: Diagram depicting nested rectangles of varying color. The outermost one titled "Grid", inside sets of uniform rectangles layered on one another titled "Block". Each "Block" containing sets of uniform rectangles layered on one another titled "Warp". Each of the "Warp" titled rectangles filled with downward pointing arrows inside. Hierarchy of thread groups. .. _work-item: Thread (Work-item) The smallest unit of execution in the HIP programming model is a thread, also called a work-item in lower-level documentation such as AMDGPU ISA manuals and HSA specifications. Each thread represents an independent control flow with its own registers and program counter. Threads execute the same kernel function independently, using identifiers such as ``threadIdx.x`` and ``blockIdx.x`` that determine which portion of data the thread operates on within its work group and grid. For example: .. code-block:: cuda int tid = threadIdx.x; int bid = blockIdx.x; int global_idx = tid + bid * blockDim.x; Each thread maintains its own register state, including vector general-purpose registers (VGPRs), and a private program counter. Threads have private storage for local variables, spilled registers, and function call stacks, which reside in global memory when on-chip register limits are exceeded. High-performance kernels minimize this usage by keeping data in registers or :ref:`Local Data Share (LDS) `. A single SIMD lane executes the instructions for one thread at a time. Because each CU contains multiple SIMD units, thousands of threads can execute concurrently across the GPU. From a programmer's perspective, a thread is the fundamental unit of GPU computation, analogous to a CPU thread but scaled to tens of thousands of instances operating in parallel. .. _wavefront: Warp (or Wavefront) The innermost grouping of threads is called a warp (HIP terminology) or wavefront (AMD ISA terminology). A warp is the most tightly coupled group of threads, both physically and logically. Threads within a warp are executed in lockstep, with each thread executing the same instruction simultaneously on different data elements. A warp represents the fundamental execution unit of AMD GPUs. Each warp consists of multiple parallel threads that execute the same instruction simultaneously across the SIMD pipelines of a :ref:`compute unit `. Threads in a warp are also called lanes, and the value identifying them is the lane ID. .. tip:: Lane IDs aren't queried like other thread IDs, but are user-calculated. As a consequence, they are only as multidimensional as the user interprets the calculated values to be. The size of a warp is architecture dependent and always fixed: * **64 threads** for CDNA architectures * **32 threads** for RDNA architectures Warps are signified by the set of communication primitives at their disposal, as discussed in :ref:`warp-cross-lane`. On modern AMD datacenter GPUs, the number of resident warps per CU is limited by architectural wave slots and available resources (such as registers and LDS). For example, if a CU supports 64 resident warps, each containing 64 threads, this would correspond to 4,096 threads in flight; actual limits are architecture-specific and described in the hardware implementation documentation. When a warp issues an instruction whose operands are not yet ready, for instance, a global memory load from :ref:`HBM `, it becomes stalled. Rather than idling, the warp scheduler selects another ready warp to execute. This rapid context switching hides memory and instruction latency and is key to achieving high utilization. To keep the compute units busy, you should maximize occupancy, the number of resident warps per CU, ensuring there is always at least one eligible warp ready to issue instructions. The ratio of active issue cycles to total cycles is known as issue efficiency. .. _inherent_thread_hierarchy_block: Block (Work-group) The next level of the thread hierarchy is called a thread block (or work-group in OpenCL terminology). A block is a collection of warps that can synchronize and share local data share (LDS) memory. The defining feature of a block is that all threads in the block have shared memory that they can use to share data or synchronize with one another, as described in :ref:`memory_hierarchy`. All warps of a block execute on the same CU, ensuring they can access the same LDS and synchronize efficiently. This locality is crucial for performance when threads need to cooperate on shared data. Threads within a work-group can coordinate with one another through barriers and the shared :ref:`Local Data Share (LDS) `. Because LDS resides in the CU's L1 cache subsystem, this cooperation is extremely fast. Threads in different work-groups cannot synchronize directly and must communicate through global memory using atomics or fences. The size of a block, or the block dimension, is the user-configurable number of threads per block, but is limited by the queryable capabilities of the executing hardware. Common limits include: * Maximum threads per block: typically 1024 * Maximum block dimensions: 1024 x 1024 x 64 (x, y, z) * Limited by available resources (registers, LDS, warp slots) The unique ID of the thread within a block can be 1, 2, or 3-dimensional as provided by the HIP API. You can configure the thread block to best represent the data associated with the kernel instruction set. .. note:: When linearizing thread IDs within a block, assume the *fast index* is the ``x`` dimension, followed by the ``y`` and ``z`` dimensions. .. _inherent_thread_hierarchy_grid: Grid The top-most level of the thread hierarchy is a grid. A grid represents the total collection of blocks (work-groups) launched for a single kernel execution. It defines the overall problem size and how work is distributed across the GPU. When a kernel is launched, the host specifies the grid's dimensions, the number of work-groups, and the number of threads per work-group. Each thread within the grid is assigned a unique index derived from its position within both its work-group and the global grid, allowing threads to cooperatively process large datasets. All work-groups in a grid execute the same kernel code, but on different portions of input data. The scheduling of work-groups is handled dynamically by the GPU driver and hardware scheduler, distributing them across available compute units. Execution order is non-deterministic, work-groups may execute concurrently or in arbitrary order depending on hardware availability and resource usage. Work-groups within a grid cannot synchronize directly through barriers, since they may execute on different CUs. Instead, they coordinate via global memory, often using atomic operations or device-wide synchronization mechanisms. The kernel itself returns only after all work-groups in the grid have finished execution. This grid-level abstraction allows a single kernel launch to scale transparently with GPU size: larger GPUs simply execute more work-groups concurrently, while smaller GPUs schedule them in more waves. The grid is specified when launching a kernel and determines: * Total number of threads: ``grid_size`` × ``block_size`` * Distribution of work across compute units * Overall parallelism of the computation The unique ID of each block within a grid can be 1, 2, or 3-dimensional, as provided by the API, and is accessible to every thread in the block through the ``blockIdx`` built-in variable. Grid dimensions are limited by queryable hardware capabilities. The three-dimensional thread hierarchy available to a kernel program lends itself to solutions that align closely to the computational problem. The following are some examples: * 1-dimensional: array processing, linear data structures, or sequential data transformation * 2-dimensional: Image processing, matrix operations, 2 dimensional simulations * 3-dimensional: Volume rendering, 3D scientific simulations, spatial algorithms Cooperative groups thread model ------------------------------- The Cooperative groups API introduces new functions to launch, group, subdivide, synchronize and identify threads, as well as some predefined group-collective algorithms. Cooperative groups let you define your own set of thread groups which may fit your use-cases better than those defined by the hardware. It relaxes some restrictions of the :ref:`inherent_thread_model` imposed by the strict 1:1 mapping of architectural details to the programming model. .. note:: The implicit groups defined by kernel launch parameters are still available when working with cooperative groups. For further information, see :doc:`Cooperative groups `. .. _memory_hierarchy: Memory model ============ The GPU memory architecture is designed to support parallel execution across the thread hierarchy. Understanding the following memory spaces and their relationships to thread groupings is crucial for efficient GPU programming. The choice of memory type and access patterns significantly impacts kernel performance. HIP's memory hierarchy consists of three main levels that mirror the GPU's physical structure: * **Private registers**: Per-thread storage with the fastest access * **Local Data Share (LDS)**: Per-work-group shared memory with low latency * **High-Bandwidth Memory (HBM)**: Device-wide global memory with high capacity This hierarchy gives you precise control over data locality and synchronization, allowing compute-bound kernels to efficiently exploit the massively parallel architecture. The following figure summarizes the memory namespaces and how they relate to the various levels of the threading model. .. figure:: ../data/understand/programming_model/memory_hierarchy.svg :alt: Diagram depicting nested rectangles of varying color. The outermost one titled "Grid", inside it are two identical rectangles titled "Block", inside them are ones titled "Local" with multiple "Warp" titled rectangles. Blocks have not just Local inside, but also rectangles titled "Shared". Inside the Grid is a rectangle titled "Global" with three others inside: "Surface", "Texture" (same color) and "Constant" (different color). Memory hierarchy. Registers or per-thread memory Read-write storage only visible to the threads defining the given variables, also called per-thread memory. This is the default memory namespace. Each thread uses registers to store temporary variables, operands, and intermediate results during execution. These registers physically reside in the register file of a :ref:`compute unit ` and are implemented using fast on-chip SRAM, typically the fastest accessible memory in the GPU. When a thread requires more storage than the available registers, some values may spill into global memory, which incurs a significant performance penalty due to much higher latency. Registers are managed automatically by the LLVM-based ROCm compiler toolchain during kernel compilation, optimizing register allocation to minimize spills and maximize the number of threads that can run concurrently on a CU. The size of the blocks for a given kernel, and thereby the number of concurrent warps, are limited by register usage. This relates to the *occupancy* of the CU as described in :doc:`Compute Units <./hardware_implementation>`, an important concept in resource usage and performance optimization. Use registers when the data is specific to a thread. This includes temporary variables, loop counters, and intermediate results that don't need to be shared across threads. Registers provide the lowest latency access for thread-specific data. Shared memory Read-write storage visible to all the threads in a given block. Shared memory corresponds to the :ref:`Local Data Share (LDS) `, a fast, on-chip SRAM region within each compute unit's L1 cache subsystem. LDS provides low-latency, programmer-managed shared memory that enables efficient coordination and reuse of data. A typical high-performance kernel follows this pattern: * Load data from global memory into LDS * Perform arithmetic operations using SIMD or :ref:`Matrix Cores (MFMA units) ` * Synchronize threads within the work-group via barrier instructions * Write results back to global memory, optionally using atomics for inter-work-group coordination Each CU provides a fixed amount of LDS (the size depending on the architecture), shared among all resident work-groups. How much LDS a kernel consumes directly influences occupancy, since larger allocations reduce the number of active work-groups per CU. Use shared memory when the data is reused within a thread block, when cross-thread communication is needed, or to minimize global memory transactions by using on-chip memory whenever possible. Global Read-write storage visible to all threads in a given grid. There are specialized versions of global memory with different usage semantics which are typically backed by the same hardware storing global. Global memory serves as the main data store for input tensors, intermediate results, and kernel outputs. On AMD GPUs, global memory is physically implemented in :ref:`High-Bandwidth Memory (HBM) ` or GDDR memory, depending on the product class. Access to global memory is explicit and programmer-managed, with synchronization possible through atomic operations or memory fences. Data transfers between the host and device are managed using HIP runtime APIs such as :cpp:func:`hipMalloc`, :cpp:func:`hipMemcpy`, and :cpp:func:`hipFree`. Performance is largely determined by memory coalescing, alignment, and reuse. Use global memory when you have large datasets, are transferring memory between the host and the device, and when you are sharing data between thread blocks. Constant Read-only storage visible to all threads in a given grid. It is a limited segment of global with queryable size. Use constant memory for read-only data that is shared across multiple threads, and that has a small data size. Texture Read-only storage visible to all threads in a given grid and accessible through additional APIs. Surface A read-write version of texture memory. Memory optimizations and best practices --------------------------------------- .. figure:: ../data/understand/programming_model/memory-access.svg :alt: Diagram depicting an example memory access pattern for coalesced memory. The diagram has uncoalesced access on the left side, with consecutive threads accessing memory in a random pattern. With coalesced access on the right showing consecutive threads accessing consecutive memory addresses. Coalesced memory accesses The following are a few memory access patterns and best practices to improve performance. You can find additional information in :ref:`memory_management` and :doc:`../how-to/performance_guidelines`. * **Global memory**: Coalescing reduces the number of memory transactions. Coalesced memory access in HIP refers to the optimization of memory transactions to maximize throughput when accessing global memory. When a kernel accesses global memory, the memory transactions typically occur in chunks of 32, 64, or 128 bytes, which must be naturally aligned. Coalescing memory accesses means aligning and organizing these accesses so that multiple threads in a warp can combine their memory requests into the fewest possible transactions. If threads access memory in a coalesced manner, meaning consecutive threads read or write consecutive memory locations, the memory controller can merge these accesses into a single transaction. This is crucial because global memory bandwidth is relatively low compared to on-chip bandwidths, and non-optimal memory accesses can significantly impact performance. If all the threads in a warp can access consecutive memory locations, memory access is fully coalesced. To achieve coalesced memory access in HIP, you should: 1. *Align Data*: Use data types that are naturally aligned and ensure that structures and arrays are aligned properly. 2. *Optimize Access Patterns*: Arrange memory accesses so that consecutive threads in a warp access consecutive memory locations. For example, if threads access a 2D array, the array and thread block widths should be multiples of the warp size. 3. *Avoid strided access*: For example array[i * stride] can lead to memory bank conflicts and inefficient access. 4. *Pad Data*: If necessary, pad data structures to ensure alignment and coalescing. * **Shared memory**: Avoiding bank conflicts reduces the serialization of memory transactions. Shared memory is a small, fast memory region inside the CU. Unlike global memory, shared memory accesses do not require coalescing, but they can suffer from bank conflicts, which are another form of inefficient memory access. Shared memory is divided into multiple memory banks (usually 32 banks on modern GPUs). If multiple threads within a warp try to access different addresses that map to the same memory bank, accesses get serialized, leading to poor performance. To optimize shared memory usage, ensure that consecutive threads access different memory banks. Use padding if necessary to avoid conflicts. * **Texture memory**: Spatial locality improves caching performance. Texture memory is read-only memory optimized for spatial locality and caching rather than coalescing. Texture memory is cached, unlike standard global memory, and it provides optimized access patterns for 2D and spatially local data. Accessing neighboring values results in cache hits, improving performance. Therefore, instead of worrying about coalescing, optimal memory access patterns involve ensuring that threads access spatially adjacent texture elements, and the memory layout aligns well with the 2D caching mechanism. * **Unified memory**: Structured access reduces the overhead of page migrations. Unified memory allows the CPU and GPU to share memory seamlessly, but performance depends on access patterns. Unified memory enables automatic page migration between CPU and GPU memory. However, if different threads access different pages, it can lead to expensive page migrations and slow throughput performance. Accessing unified memory in a structured, warp-friendly manner reduces unnecessary page transfers. Ensure threads access memory in a structured, consecutive manner, minimizing page faults. Prefetch data to the GPU before computation by using ``hipMemPrefetchAsync()``. In addition, using small batch transfers as described below, can reduce unexpected page migrations when using unified memory. * **Small batch transfers**: Enable pipelining and improve PCIe bandwidth use. Memory transfers between the host and the device can become a major bottleneck if not optimized. One method is to use small batch memory transfers where data is transferred in smaller chunks instead of dealing with large datasets to avoid long blocking operations. Small batch transfers offer better PCIe bandwidth utilization over large data transfers. Small batch transfers offer performance improvement by offering reduced latency with small batches that run asynchronously using ``hipMemcpyAsync()`` as described in :ref:`asynchronous_how-to`, pipelining data transfers and kernel execution using separate streams. Finally, using pinned memory with small batch transfers enables faster DMA transfers without CPU involvement, greatly improving memory transfer performance. Execution model =============== As previously discussed in :ref:`heterogeneous_programming`, HIP programs consist of two distinct scopes: * The host-side API running on the host processor. * The device-side kernels running on GPUs. Both the host and the device-side APIs have synchronous and asynchronous functions. Host-side execution ------------------- The host-side API dealing with device management and their queries are synchronous. All asynchronous APIs, such as kernel execution, data movement and potentially data allocation/freeing all happen in the context of device streams, as described in `Managing streams <../how-to/hip_runtime_api/asynchronous.html#managing-streams>`_. Streams are FIFO buffers of commands to execute relating to a given device. Operations that enqueue tasks on a stream all return promptly, and the command is executed asynchronously. All side effects of a command on a stream are visible to all subsequent commands on the same stream. Multiple streams may point to the same device and those streams may be fed from multiple concurrent host-side threads. Execution on multiple streams may be concurrent but isn't required to be. Asynchronous APIs involving a stream all return a stream event, which can be used to synchronize the execution of multiple streams. A user may enqueue a barrier onto a stream referencing an event. The barrier will block activity on the stream until the operation related to the event completes. After the event completes, all side effects of the operation will be visible to subsequent commands even if those side effects manifest on different devices. .. figure:: ../data/understand/programming_model/stream-workflow.svg :alt: Diagram depicting the stream and event workflow, with an example of multiple streams working together. The diagram shows operations as red rectangles, and events as white dots. There are three streams labelled Stream 1, 2, and 3. The streams each have multiple operations and events that require synchronization between the streams. Multiple stream workflow Streams also support executing user-defined functions as callbacks on the host. The stream will not launch subsequent commands until the callback completes. Device-side execution --------------------- Kernels may be launched in multiple ways, all with different syntaxes and intended use cases. * Using the triple-chevron ``<<<...>>>`` operator on a ``__global__`` annotated function. * Using ``hipLaunchKernelGGL()`` on a ``__global__`` annotated function. .. tip:: This name, by default, is a macro expanding to the triple-chevron syntax. In cases where language syntax extensions are undesirable, or where launching templated and/or overloaded kernel functions define the ``HIP_TEMPLATE_KERNEL_LAUNCH`` preprocessor macro before including the HIP headers to turn it into a templated function. Asynchronous execution ---------------------- Asynchronous operations between the host and the kernel provide a variety of opportunities, or challenges, for managing synchronization, as described in :ref:`asynchronous_how-to`. For instance, a basic model would be to launch an asynchronous operation on a kernel in a stream, create an event to track the operation, continue operations in the host program, and when the event shows that the asynchronous operation is complete, synchronize the kernel to return the results. However, one of the opportunities of asynchronous operation is the pipelining of operations between launching kernels and transferring memory. In this case, you would be working with multiple streams running concurrently, or at least overlapping in some regard, and managing any dependencies between the streams in the host application. The producer-consumer paradigm can be used to convert a sequential program into parallel operations to improve performance. This process can employ multiple streams to kick off asynchronous kernels, provide data to the kernels, perform operations, and return the results for further processing in the host application. These asynchronous activities call for stream management strategies. In the case of the single stream, the only management would be the stream synchronization when the work was complete. However, with multiple streams you have overlapping execution of operations and synchronization becomes more complex, as shown in the variations of the example in `Programmatic dependent launch and synchronization <../how-to/hip_runtime_api/asynchronous.html#programmatic-dependent-launch-and-synchronization>`_. You need to manage each stream's activities, evaluate the availability of results, evaluate the critical path of the tasks, allocate resources on the hardware, and manage the execution order. Multi-GPU and load balancing ---------------------------- For applications requiring additional computational power beyond a single device, HIP supports utilizing multiple GPUs within a system. Large-scale applications that need more compute power can use multiple GPUs in the system. This enables the runtime to distribute workloads across multiple GPUs to balance the load and prevent some GPUs from being over-utilized while others are idle. For more information, see :ref:`multi-device`. Domain-specific programming models ================================== Beyond the low-level kernel-based programming model described in previous sections, ROCm provides domain-specific library abstractions that offer alternative ways to express GPU computation. These libraries encapsulate common computational patterns, providing higher-level programming models optimized for specific problem domains. Rather than writing explicit kernels with manual memory management and optimization, developers can express algorithms using domain-appropriate operations. The libraries handle kernel selection, memory layout, and performance tuning automatically, adapting to different GPU architectures and problem sizes at runtime. This section presents two representative examples: :doc:`rocBLAS ` for linear algebra and :doc:`MIOpen ` for deep learning. These libraries illustrate how domain-specific abstractions can improve productivity while maintaining performance portability across AMD GPU architectures. Linear algebra with rocBLAS ---------------------------- rocBLAS implements the BLAS standard, providing a high-performance library for fundamental dense linear algebra operations. It represents an alternative programming model where computation is expressed as matrix and vector operations rather than explicit GPU kernels. Programming model characteristics ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The rocBLAS programming model differs from kernel-level HIP in several key aspects: * **Abstraction level**: Developers call functions like ``rocblas_sgemm()`` for matrix multiplication instead of writing tiled kernels manually * **Memory layout**: Matrices are conceptualized as mathematical objects with rows and columns, not raw pointers and index arithmetic * **Automatic optimization**: The library selects optimal kernel implementations based on problem dimensions, data types (FP64, FP32, FP16, BF16, INT8), and target GPU architecture (for example, ``gfx90a``, ``gfx942``) * **Performance portability**: The same API call adapts to CDNA, RDNA, and future architectures without code changes Instead of managing thread blocks, shared memory tiles, and register allocation, the developer specifies the mathematical operation. rocBLAS handles all low-level optimizations, including memory coalescing, use of :ref:`Matrix Cores (MFMA units) `, and exploitation of on-chip :ref:`LDS `. Column-major versus row-major layouts ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ One practical consideration when using rocBLAS is matrix layout. To maintain compatibility with the original Fortran-based BLAS specification, rocBLAS expects matrices in column-major order. Most C and C++ programs, including HIP kernels, use row-major order by default. This mismatch can be resolved mathematically without physically transposing data. Given the identity: .. math:: \text{If } \pmb{C} = \pmb{A} \times \pmb{B}, \text{ then } \pmb{C}^T = \pmb{B}^T \times \pmb{A}^T By swapping the operand order and dimensions when calling rocBLAS, and interpreting the result as a row-major matrix, the correct result is obtained without memory rearrangement: .. code-block:: cpp #include // Matrix multiply C = alpha * A @ B + beta * C // for row-major matrices using rocBLAS void sgemm_row_major(rocblas_handle handle, int M, int N, int K, const float* alpha, const float* A, const float* B, const float* beta, float* C) { // Swap A and B, swap M and N for row-major layout rocblas_sgemm(handle, rocblas_operation_none, rocblas_operation_none, N, M, K, alpha, B, N, // leading dimension of B^T A, K, // leading dimension of A^T beta, C, N); // leading dimension of C^T } This technique preserves both correctness and performance. The ``rocblas_operation_none`` flag indicates matrices should be used as provided, avoiding additional device-side transposes that would harm performance. Integration with HIP applications ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ rocBLAS integrates with HIP applications through a handle-based API. The library manages its own stream-based execution model, coordinating with HIP's asynchronous operations: .. code-block:: cpp rocblas_handle handle; rocblas_create_handle(&handle); // Optional: associate with HIP stream rocblas_set_stream(handle, hipStream); // Perform computation rocblas_sgemm(handle, ...); // Synchronize if needed hipStreamSynchronize(hipStream); rocblas_destroy_handle(handle); Applications link against ``librocblas.so`` and include ``rocblas.h``. The library serves as the foundation for deep learning frameworks such as PyTorch and TensorFlow when running on AMD GPUs, handling dense matrix operations in fully connected and transformer layers. Deep learning with MIOpen -------------------------- MIOpen provides GPU-accelerated primitives for deep neural networks, offering a programming model centered on neural network operations rather than explicit kernels. It occupies the same software layer as rocBLAS but is specialized for deep learning workloads. Programming model characteristics ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ MIOpen abstracts GPU programming to the level of neural network layers and operations: * **Domain expressiveness**: Computations are expressed as convolutions, normalizations, activations, and attention mechanisms, not thread hierarchies * **Declarative fusion**: Operations can be combined (Convolution → Bias → ReLU) into fused kernels automatically * **Automatic algorithm selection**: The library chooses between direct, FFT-based, Winograd, and GEMM-based implementations at runtime * **Memory optimization**: Fused operations keep intermediate results in :ref:`LDS `, minimizing :ref:`global memory ` traffic Rather than manually implementing convolutional kernels with careful attention to memory coalescing and :ref:`MFMA unit ` utilization, developers describe the neural network architecture. MIOpen handles low-level optimization, auto-tuning, and performance-critical implementation details. Graph and Fusion APIs ~~~~~~~~~~~~~~~~~~~~~~ Modern MIOpen workflows use Graph and Fusion APIs, which allow declarative specification of operation sequences: .. code-block:: cpp // Declarative fusion example (conceptual) auto conv = createConvolution(input, weights, ...); auto bias = createBiasAdd(conv, bias_vector); auto relu = createActivation(bias, ACTIVATION_RELU); // Library fuses into single optimized kernel auto fused_op = fuseOperations({conv, bias, relu}); executeGraph(fused_op); Operation fusion significantly improves performance by reducing memory bandwidth pressure. Instead of three separate kernel launches with intermediate global memory writes and reads, a single fused kernel keeps data on-chip throughout the computation pipeline. This is particularly important for operations like batch normalization followed by activation, where intermediate tensors would otherwise require expensive round-trips through :ref:`HBM `. Keeping these values in LDS reduces memory traffic and improves energy efficiency. Runtime optimization and tuning ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ MIOpen maintains multiple implementations for each operation. For convolutions, this includes: * **Direct convolution**: Explicit loops over kernel dimensions * **GEMM-based**: Reformulates convolution as matrix multiplication * **Winograd**: Reduces arithmetic complexity for small kernels * **FFT-based**: Frequency-domain convolution for large kernels The library uses runtime heuristics and pre-tuned configuration databases to select the optimal algorithm based on: * Input tensor dimensions and layout * Filter sizes and stride patterns * Target GPU architecture (CDNA versus RDNA, specific ``gfx`` version) * Available memory and compute resources * Data types (FP32, FP16, BF16, INT8) This auto-tuning abstracts performance optimization from the application developer. The same MIOpen call automatically adapts when moving from a datacenter MI300X (``gfx942``) to a consumer RDNA3 GPU (``gfx1200``), selecting architecture-appropriate implementations without code changes. Framework integration ~~~~~~~~~~~~~~~~~~~~~ Deep learning frameworks like PyTorch and TensorFlow use MIOpen as their GPU backend on AMD hardware. When a PyTorch model calls ``torch.nn.Conv2d`` or ``torch.nn.BatchNorm2d``, these operations dispatch to MIOpen primitives. The framework provides the high-level neural network API, while MIOpen handles the GPU-specific implementation. This separation allows framework developers to focus on model architectures and training algorithms, while library developers optimize GPU kernels for specific hardware generations. For applications requiring both general linear algebra and specialized deep learning operations, rocBLAS and MIOpen work together. MIOpen uses rocBLAS (via hipBLASLt and Tensile backends) for the matrix-multiply-intensive portions of operations like fully connected layers, while providing specialized kernels for convolutions, pooling, and normalization. Comparison of programming models --------------------------------- The following table summarizes how different programming models trade control for abstraction: .. list-table:: Programming Model Comparison :header-rows: 1 :widths: 30 35 35 * - Aspect - HIP Kernels - Domain-Specific Libraries * - Programming unit - Thread, block, grid - Domain primitives (matrices, layers, transforms) * - Memory management - Explicit (``hipMalloc``, ``hipMemcpy``) - Library-managed, handle-based * - Optimization approach - Manual tuning required - Auto-tuned, architecture-adaptive * - Thread hierarchy - Explicit configuration - Abstracted by API * - Performance portability - Requires retuning per architecture - Automatic adaptation across architectures * - Development effort - High (low-level control) - Low (high-level API calls) The examples of rocBLAS and MIOpen illustrate a common pattern across the ROCm software stack: domain-specific libraries that abstract low-level GPU programming into higher-level operations tailored to specific computational domains. The ROCm ecosystem provides many additional libraries following this pattern, including rocFFT for Fast Fourier Transforms, rocRAND for random number generation, rocSPARSE for sparse linear algebra, and RCCL for collective communication across multiple GPUs. Choosing the appropriate abstraction level depends on the application domain. For novel algorithms not covered by existing libraries, kernel-level HIP provides full control. For well-established operations within specific domains, the ROCm library ecosystem offers productivity and performance portability. Many applications combine multiple programming models: HIP kernels for custom algorithms, domain-specific libraries for standard operations, and framework integration for high-level workflows. The ROCm software stack supports this layered approach, allowing developers to select the appropriate abstraction for each component of their application. For a complete list of ROCm libraries and their capabilities, see the `ROCm documentation `_.