> For the complete Mojo documentation index, see [llms.txt](/llms.txt).
> Markdown versions of all pages are available by appending .md to any URL (e.g. /docs/manual/basics.md).
# GPU programming fundamentals
This guide explores the fundamentals of GPU programming using the Mojo
programming language, covering essential concepts and techniques for developing
GPU-accelerated applications that can work on a variety of supported GPUs from
different vendors.
Key topics covered in this guide:
- Understanding the CPU-GPU programming model.
- Working with Mojo's GPU support through the standard library.
- Managing GPU devices and contexts using `DeviceContext`.
- Writing and executing kernel functions for parallel computation.
- Memory management and data transfer between CPU and GPU.
- Organizing threads and thread blocks for optimal performance.
Before diving into GPU programming, ensure you have a [compatible
GPU](/docs/requirements#gpu-compatibility) and the necessary development
environment installed. And if you're new to GPU programming, we suggest you
read the [Intro to GPU architectures](/docs/manual/gpu/architecture/).
## Overview of GPU programming in Mojo
The Mojo language, including its standard library and open source MAX kernels
library, allow you to develop GPU-enabled applications.
### GPU support in the Mojo standard library
The [`gpu`](/docs/std/gpu/) package of the Mojo standard library includes
several subpackages for interacting with GPUs, with the
[`gpu.host`](/docs/std/gpu/host/) package providing most of the commonly used
APIs. However, the [`sys`](/docs/std/sys/info/) package contains a few basic
introspection functions for determining whether a system has a supported GPU:
- [`has_accelerator()`](/docs/std/sys/info/has_accelerator/): Returns `True`
if the host system has an accelerator and `False` otherwise.
- [`has_amd_gpu_accelerator()`](/docs/std/sys/info/has_amd_gpu_accelerator/):
Returns `True` if the host system has an AMD GPU and `False` otherwise.
- [`has_apple_gpu_accelerator()`](/docs/std/sys/info/has_apple_gpu_accelerator/):
Returns `True` if the host system has an Apple silicon GPU and `False`
otherwise.
- [`has_nvidia_gpu_accelerator()`](/docs/std/sys/info/has_nvidia_gpu_accelerator/):
Returns `True` if the host system has an NVIDIA GPU and `False` otherwise.
These functions are useful for conditional compilation or execution depending on
whether a supported GPU is available.
```mojo title="detect_gpu.mojo"
from std.sys import has_accelerator
def main():
comptime if has_accelerator():
print("GPU detected")
# Enable GPU processing
else:
print("No GPU detected")
# Print error or fall back to CPU-only execution
```
:::note
Mojo requires a
[compatible GPU development environment](/docs/requirements/#gpu-compatibility)
to compile kernel functions, otherwise it raises a compile-time error. In this
example, we're using the
[`comptime if`](/docs/manual/metaprogramming/comptime-evaluation/#comptime-if)
statement to evaluate the `has_accelerator()` function at compile time and
compile only the corresponding branch of the `if` statement. As a result, if you
don't have a compatible GPU development environment, you'll see the following
message when you run the program:
```output
No GPU detected
```
:::
### GPU programming model
GPU programming follows a distinct pattern where work is divided between the CPU
and GPU:
- The CPU (host) manages program flow and coordinates GPU operations.
- The GPU (device) executes parallel computations across many threads.
- You must explicitly manage data exchange between host and device memory.
A GPU program generally follows these steps:
1. Initialize data in host (CPU) memory.
2. Allocate device (GPU) memory and transfer data from host to device memory.
3. Execute a kernel function on the GPU to process the data.
4. Transfer results back from device to host memory.
This process typically runs asynchronously, allowing the CPU to perform other
tasks while the GPU processes data. Any time that the CPU needs to ensure that
the GPU has completed an operation, such as before it copies kernel results from
device memory, it must first explicitly synchronize with the GPU as described in
[Asynchronous operation and synchronizing the CPU and
GPU](#asynchronous-operation-and-synchronizing-the-cpu-and-gpu).
A simple example helps to understand this programming model. We'll not go into
detail about the specific APIs at this point other than the included comments,
but all of the types, functions, and methods are discussed in more detail in
later sections of this document.
```mojo title="scalar_add.mojo"
from std.math import iota
from std.sys import exit, has_accelerator
from std.gpu.host import DeviceContext
from std.gpu import block_dim, block_idx, thread_idx
comptime num_elements = 20
def scalar_add(
vector: UnsafePointer[Float32, MutAnyOrigin], size: Int, scalar: Float32
):
"""
Kernel function to add a scalar to all elements of a vector.
This kernel function adds a scalar value to each element of a vector stored
in GPU memory. The input vector is modified in place.
Args:
vector: Pointer to the input vector.
size: Number of elements in the vector.
scalar: Scalar to add to the vector.
"""
# Calculate the global thread index within the entire grid. Each thread
# processes one element of the vector.
#
# block_idx.x: index of the current thread block.
# block_dim.x: number of threads per block.
# thread_idx.x: index of the current thread within its block.
idx = block_idx.x * block_dim.x + thread_idx.x
# Bounds checking: ensure we don't access memory beyond the vector size.
# This is crucial when the number of threads doesn't exactly match vector
# size.
if idx < size:
# Each thread adds the scalar to its corresponding vector element
# This operation happens in parallel across all GPU threads
vector[idx] += scalar
def main() raises:
comptime if not has_accelerator():
print("No GPUs detected")
exit(0)
else:
# Initialize GPU context for device 0 (default GPU device).
ctx = DeviceContext()
# Create a buffer in host (CPU) memory to store our input data
host_buffer = ctx.enqueue_create_host_buffer[DType.float32](https://mojolang.org/docs/manual/gpu/num_elements.md)
# Wait for buffer creation to complete.
ctx.synchronize()
# Fill the host buffer with sequential numbers (0, 1, 2, ..., size-1).
iota(host_buffer.as_span())
print("Original host buffer:", host_buffer)
# Create a buffer in device (GPU) memory to store data for computation.
device_buffer = ctx.enqueue_create_buffer[DType.float32](https://mojolang.org/docs/manual/gpu/num_elements.md)
# Copy data from host memory to device memory for GPU processing.
ctx.enqueue_copy(src_buf=host_buffer, dst_buf=device_buffer)
# Compile the scalar_add kernel function for execution on the GPU.
scalar_add_kernel = ctx.compile_function[scalar_add]()
# Launch the GPU kernel with the following arguments:
#
# - device_buffer: GPU memory containing our vector data
# - num_elements: number of elements in the vector
# - Float32(20.0): the scalar value to add to each element
# - grid_dim=1: use 1 thread block
# - block_dim=num_elements: use 'num_elements' threads per block (one
# thread per vector element)
ctx.enqueue_function(
scalar_add_kernel,
device_buffer,
num_elements,
Float32(20.0),
grid_dim=1,
block_dim=num_elements,
)
# Copy the computed results back from device memory to host memory.
ctx.enqueue_copy(src_buf=device_buffer, dst_buf=host_buffer)
# Wait for all GPU operations to complete.
ctx.synchronize()
# Display the final results after GPU computation.
print("Modified host buffer:", host_buffer)
```
This application produces the following output:
```output
Original host buffer: HostBuffer([0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0, 19.0])
Modified host buffer: HostBuffer([20.0, 21.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0, 28.0, 29.0, 30.0, 31.0, 32.0, 33.0, 34.0, 35.0, 36.0, 37.0, 38.0, 39.0])
```
## Accessing and managing GPUs with `DeviceContext`
The [`gpu.host`](/docs/std/gpu/host/) package includes the
[`DeviceContext`](/docs/std/gpu/host/device_context/DeviceContext/) struct,
which represents a logical instance of a GPU device. It provides methods for
allocating memory on the device, copying data between the host CPU and the GPU,
and compiling and running functions (also known as *kernels*) on the device.
### Creating an instance of `DeviceContext` to access a GPU
Mojo supports systems with multiple GPUs. GPUs are uniquely identified by
integer indices starting with `0`, which is considered the "default" device. You
can determine the number of GPUs available by invoking the
[`DeviceContext.number_of_devices()`](/docs/std/gpu/host/device_context/DeviceContext/#number_of_devices)
static method.
The `DeviceContext()` constructor returns an instance for interacting with a
specified GPU. It accepts two optional arguments:
- `device_id`: An integer index of a specific GPU on the system. The default
value of 0 refers to the "default" GPU for the system.
- `api`: A `String` specifying a particular vendor's API. "cuda" (NVIDIA), "hip"
(AMD), and "metal" (Apple) are currently supported.
If your system doesn't have a supported GPU—or doesn't have a GPU matching the
`device_id` or `api`, if provided—then the constructor raises an error.
### Asynchronous operation and synchronizing the CPU and GPU
Typical CPU-GPU interaction is asynchronous, allowing the GPU to process tasks
while the CPU is busy with other work. Each `DeviceContext` has an associated
stream of queued operations to execute on the GPU. Operations within a stream
execute in the order they are enqueued.
The
[`synchronize()`](/docs/std/gpu/host/device_context/DeviceContext/#synchronize)
method blocks execution of the current CPU thread until all queued operations on
the associated `DeviceContext` stream have completed. Most commonly, you use
this to wait until the result of a kernel function is copied from device memory
to host memory before accessing it on the host.
## Kernel functions
A GPU *kernel* is simply a function that runs on a GPU, executing a specific
computation on a large dataset in parallel across thousands or millions of
*threads*. You specify the number of threads when you execute a kernel function,
and all threads run the same kernel function. However, the GPU assigns a unique
thread index for each thread, and you use the thread index to determine which
data elements an individual thread should process.
### Multidimensional grids and thread organization
As discussed in [GPU execution
model](/docs/manual/gpu/architecture#gpu-execution-model), a *grid* is the
top-level organizational structure of the threads executing a kernel function on
a GPU. A grid consists of multiple *thread blocks*, which are organized across
one, two, or three dimensions. Each thread block is further divided into
individual threads, which are in turn organized across one, two, or three
dimensions.
You specify the grid and thread block dimensions with the `grid_dim` and
`block_dim` keyword arguments when you enqueue a kernel function to execute
using the
[`enqueue_function()`](/docs/std/gpu/host/device_context/DeviceContext/#enqueue_function)
method. For example:
```mojo
# Enqueue the print_threads() kernel function
ctx.enqueue_function[print_threads](https://mojolang.org/docs/manual/gpu/grid_dim=(2, 2, 1.md), # 2x2x1 blocks per grid
block_dim=(4, 4, 2), # 4x4x2 threads per block
)
```
For both `grid_dim` and `block_dim`, you express the size in the `x`, `y`, and
`z` dimensions as a [`Dim`](/docs/std/gpu/host/dim/Dim/) or a `Tuple`. The
`y` and `z` dimensions default to 1 if you don't explicitly provide them (that
is, `(2, 2)` is treated as `(2, 2, 1)` and `(8,)` is treated as `(8, 1, 1)`).
You can also provide just an `Int` value to specify only the `x` dimension (that
is, `64` is treated as `(64, 1, 1)`).
**Figure 1.** Organization of thread blocks and threads within a
grid.
From within a kernel function, you can access the grid and thread block
dimensions and the assigned thread block and thread indices of the individual
threads executing the kernel using the following structures:
| Comptime value | Description |
|---------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [`grid_dim`](/docs/std/gpu/primitives/id/#grid_dim) | Dimensions of the grid as `x`, `y`, and `z` values (for example, `grid_dim.y`). |
| [`block_dim`](/docs/std/gpu/primitives/id/#block_dim) | Dimensions of the thread block as `x`, `y`, and `z` values. |
| [`block_idx`](/docs/std/gpu/primitives/id/#block_idx) | Index of the block within the grid as `x`, `y`, and `z` values. |
| [`thread_idx`](/docs/std/gpu/primitives/id/#thread_idx) | Index of the thread within the block as `x`, `y`, and `z` values. |
| [`global_idx`](/docs/std/gpu/primitives/id/#global_idx) | The global offset of the thread as `x`, `y`, and `z` values. That is, `global_idx.x = block_dim.x * block_idx.x + thread_idx.x`, `global_idx.y = block_dim.y * block_idx.y + thread_idx.y`, and `global_idx.z = block_dim.z * block_idx.z + thread_idx.z`. |
:::note
All of these dimensions and indices are
[`Int`](/docs/std/builtin/int/Int/) values.
:::
The following example shows a kernel function that prints the
thread block index, thread index, and global index for each thread.
```mojo title="print_threads.mojo"
from std.sys import has_accelerator
from std.gpu.host import DeviceContext
from std.gpu import block_dim, block_idx, global_idx, thread_idx
def print_threads():
"""Print thread block and thread indices."""
print(
block_idx.x,
block_idx.y,
block_idx.z,
thread_idx.x,
thread_idx.y,
thread_idx.z,
global_idx.x,
global_idx.y,
global_idx.z,
block_dim.x * block_idx.x + thread_idx.x,
block_dim.y * block_idx.y + thread_idx.y,
block_dim.z * block_idx.z + thread_idx.z,
sep="\t",
)
def main() raises:
comptime if not has_accelerator():
print("No compatible GPU found")
else:
# Initialize GPU context for device 0 (default GPU device).
ctx = DeviceContext()
print("block_idx\t\tthread_idx\t\tglobal_idx\t\tcalculated global_idx")
print("x\ty\tz", "x\ty\tz", "x\ty\tz", "x\ty\tz", sep="\t")
print("-" * 20, "-" * 20, "-" * 20, "-" * 20, sep="\t")
ctx.enqueue_function[print_threads](https://mojolang.org/docs/manual/gpu/grid_dim=(2, 2, 1.md), # 2x2x1 blocks per grid
block_dim=(4, 4, 2), # 4x4x2 threads per block
)
ctx.synchronize()
print("Done")
```
This example produces output similar to the following (with the output order
indeterminate because of the concurrent execution of multiple threads):
```output
block_idx thread_idx global_idx calculated global_idx
x y z x y z x y z x y z
-------------------- -------------------- -------------------- --------------------
0 1 0 0 0 0 0 4 0 0 4 0
0 1 0 1 0 0 1 4 0 1 4 0
...
1 1 0 2 3 1 6 7 1 6 7 1
1 1 0 3 3 1 7 7 1 7 7 1
Done
```
:::note Known limitation
On Apple silicon GPUs, each `print()` call inside a GPU kernel currently
supports at most one string literal argument. To ensure readable output across
all GPU platforms, the examples here print numeric values only, with the host
printing column headers before launching the kernel. This restriction doesn't
apply to NVIDIA or AMD GPUs.
:::
### Writing a kernel function
Kernel functions must be
[non-raising](/docs/manual/functions/#raising-and-non-raising-functions).
:::note
For functions called from GPU kernels that need error handling, Mojo supports
typed errors with types that don't allocate heap memory. See
[Define a custom error type](/docs/manual/errors/#define-a-custom-error-type)
for details.
:::
Argument values must be of types that conform to the
[`DevicePassable`](/docs/std/builtin/device_passable/DevicePassable/) trait.
Additionally, a kernel function can't have a return value. Instead, you must
write any result of a kernel function to a memory buffer passed in as an
argument. The next two sections, [Passing data between CPU and
GPU](#passing-data-between-cpu-and-gpu) and
[`DeviceBuffer` and `HostBuffer`](#devicebuffer-and-hostbuffer) go into more
detail on how to pass values to a kernel function and get back results.
As discussed in [GPU execution
model](/docs/manual/gpu/architecture#gpu-execution-model), when the GPU executes
a kernel, it assigns the grid's thread blocks to various streaming
multiprocessors (SMs) for execution. The SM then divides the thread block into
subsets of threads called a *warp*. The size of a warp depends on the GPU
architecture, but most modern GPUs currently use a warp size of 32 or 64
threads.
**Figure 2.** Hierarchy of threads running on a GPU, showing the
relationship of the grid, thread blocks, warps, and individual threads, based
on HIP Programming Guide ©2023-2025 Advanced Micro Devices,
Inc.
If a thread block contains a number of threads not evenly divisible by the warp
size, the SM creates a partially filled final warp that still consumes the full
warp's resources. For example, if a thread block has 100 threads and the warp
size is 32, the SM creates:
- 3 full warps of 32 threads each (96 threads total).
- 1 partial warp with only 4 active threads but still occupying a full warp's
worth of resources (32 thread slots).
Because of this execution model, you must ensure that the threads in your kernel
don't attempt to access out-of-bounds data. Otherwise, your kernel might crash
or produce incorrect results. For example, if you pass a 2,000-element vector to
a kernel that you execute with single-dimension thread blocks of 512 threads
each, and each thread is responsible for processing one element, your kernel
could perform a boundary check like this to ensure that it doesn't attempt to
process out-of-bounds elements:
```mojo
from std.gpu import global_idx
def process_vector(vector: UnsafePointer[Float32, MutAnyOrigin], size: Int):
if global_idx.x < size:
# Process vector[global_idx.x] in some way
```
### Passing data between CPU and GPU
All values passed to a kernel function must be of types that conform to the
[`DevicePassable`](/docs/std/builtin/device_passable/DevicePassable/) trait.
The trait declares a
[`comptime` member](/docs/manual/traits/#comptime-members-for-generics) named
`device_type` that maps the type as used on the CPU host to a corresponding type
used on the GPU device.
As an example,
[`DeviceBuffer`](/docs/std/gpu/host/device_context/DeviceBuffer/) is a
host-side representation of a buffer located in the GPU's global memory space.
But it defines its `device_type` member as `UnsafePointer`, so the
data represented by a `DeviceBuffer` is actually passed to the kernel function
as a value of type
[`UnsafePointer`](/docs/std/memory/unsafe_pointer/UnsafePointer/). The next
section, [`DeviceBuffer` and `HostBuffer`](#devicebuffer-and-hostbuffer),
describes in more detail how to allocate memory buffers on the host and device
and to exchange blocks of data between host and device.
The following table lists the most commonly used types in the Mojo libraries
that conform to the `DevicePassable` trait.
| Host type | Device Type | Description |
|--------------------------------------------------------------------------|-----------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------|
| [`Int`](/docs/std/builtin/int/Int/) | `Int` | Signed integer |
| [`SIMD[dtype, width]`](/docs/std/builtin/simd/SIMD/) | `SIMD[dtype, width]` | Small vector backed by a hardware vector element |
| [`DeviceBuffer[dtype]`](/docs/std/gpu/host/device_context/DeviceBuffer/) | [`UnsafePointer[SIMD[dtype, 1]]`](/docs/std/memory/unsafe_pointer/UnsafePointer/) | Memory buffer of `dtype` values |
| [`TileTensor`](/docs/layout/tile_tensor/TileTensor/) | `TileTensor` | A powerful abstraction for multi-dimensional data. See [Using `TileTensor`](/docs/manual/tile-tensor/tensors/). |
### `DeviceBuffer` and `HostBuffer`
This section describes how to use `DeviceBuffer` and `HostBuffer` to allocate
memory on the device and host respectively, and to copy data between device and
host memory.
#### Creating a `DeviceBuffer`
The
[`DeviceBuffer`](/docs/std/gpu/host/device_context/DeviceBuffer/)
type represents a block of device memory associated with a particular
`DeviceContext`. Specifically, the buffer is located in the device's *global
memory* space. As such, the buffer is accessible by all threads of all kernel
functions executed by the `DeviceContext`.
As discussed in [Passing data between CPU and
GPU](#passing-data-between-cpu-and-gpu), `DeviceBuffer` is the type used by the
**host** to allocate the buffer and to copy data between the host and device.
But when you pass a `DeviceBuffer` to a kernel function, the argument received
by the function is of type `UnsafePointer`. Attempting to use the `DeviceBuffer`
type directly from within a kernel function results in an error.
The
[`DeviceContext.enqueue_create_buffer()`](/docs/std/gpu/host/device_context/DeviceContext/#enqueue_create_buffer)
method creates a `DeviceBuffer` associated with that `DeviceContext`. It accepts
the data type as a compile-time [`DType`](/docs/std/builtin/dtype/DType/)
parameter and the size of the buffer as a run-time argument. So to create a
buffer for 1,024 [`Float32`](/docs/std/builtin/simd/#float32) values, you would
execute:
```mojo
device_buffer = ctx.enqueue_create_buffer[DType.float32](https://mojolang.org/docs/manual/gpu/1024.md)
```
As the method name implies, this method is asynchronous and enqueues the
operation on the `DeviceContext`'s associated [stream of queued
operations](#asynchronous-operation-and-synchronizing-the-cpu-and-gpu).
#### Creating a `HostBuffer`
The [`HostBuffer`](/docs/std/gpu/host/device_context/HostBuffer/) type is
analogous to `DeviceBuffer`, but represents a block of host memory associated
with a particular `DeviceContext`. It supports methods for transferring data
between host and device memory, as well as a basic set of methods for accessing
data elements by index and for printing the buffer.
The
[`DeviceContext.enqueue_create_host_buffer()`](/docs/std/gpu/host/device_context/DeviceContext/#enqueue_create_host_buffer)
method accepts the data type as a compile-time
[`DType`](/docs/std/builtin/dtype/DType/) parameter and the size of the buffer
as a run-time argument and returns a `HostBuffer`. As with all `DeviceContext`
methods whose name starts with `enqueue_`, the method is asynchronous and
returns immediately, adding the operation to the queue to be executed by the
`DeviceContext`. Therefore, you need to call the `synchronize()` method to
ensure that the operation has completed before you write to or read from the
`HostBuffer` object.
```mojo
device_buffer = ctx.enqueue_create_host_buffer[DType.float32](https://mojolang.org/docs/manual/gpu/1024.md)
# Synchronize to wait until buffer is created before attempting to write to it
ctx.synchronize()
# Now it's safe to write to the buffer
for i in range(1024):
device_buffer[i] = Float32(i * i)
```
#### Copying data between host and device memory
The
[`enqueue_copy()`](/docs/std/gpu/host/device_context/DeviceContext/#enqueue_copy)
method is overloaded to support copying from host to device, device to host, or
even device to device for systems that have multiple GPUs. Typically, you'll use
it to copy data that you've staged in a `HostBuffer` to a `DeviceBuffer` before
executing a kernel, and then from a `DeviceBuffer` to a `HostBuffer` to retrieve
the results of kernel execution. The `scalar_add.mojo` example in
[GPU programming model](#gpu-programming-model) shows this pattern in action. In
it, the kernel function does an in-place modification of the buffer it receives
as an argument and then reuses the original `HostBuffer` to copy the results
back from the device. However, you can allocate a separate `DeviceBuffer` and
`HostBuffer` for the result of a kernel function if you want to retain the
original data.
In addition to copying data between a `HostBuffer` to a `DeviceBuffer`, you can
use an [`UnsafePointer`](/docs/std/memory/unsafe_pointer/UnsafePointer/) as
the source or destination of a copy. However, the `UnsafePointer` must reference
host memory for this operation. Attempting to use an `UnsafePointer` referencing
device memory results in an error. For example, this is useful if you have data
already staged in a data structure on the host that can expose the data through
an `UnsafePointer`. In that case you would not need to copy the data from the
data structure to a `HostBuffer` before copying it to the `DeviceBuffer`.
Both `DeviceBuffer` and `HostBuffer` also include
[`enqueue_copy_to()`](/docs/std/gpu/host/device_context/DeviceBuffer/#enqueue_copy_to)
and
[`enqueue_copy_from()`](/docs/std/gpu/host/device_context/DeviceBuffer/#enqueue_copy_from)
methods. These are simply convenience methods that call the `enqueue_copy()`
method on their corresponding `DeviceContext`. For example, the following two
method calls are interchangeable:
```mojo
ctx.enqueue_copy(src_buf=host_buffer, dst_buf=device_buffer)
# Equivalent to:
host_buffer.enqueue_copy_to(dst=device_buffer)
```
Finally, as a convenience for testing or prototyping, you can use the
[`DeviceBuffer.map_to_host()`](/docs/std/gpu/host/device_context/DeviceBuffer/#map_to_host)
method to create a host-accessible view of the device buffer's contents. This
returns `HostBuffer` as a
[context manager](/docs/manual/errors/#use-a-context-manager) that contains a
copy of the data from the corresponding `DeviceBuffer`. Additionally, any
modifications that you make to the `HostBuffer` are automatically copied back to
the `DeviceBuffer` when the `with` statement exits. For example:
```mojo
ctx = DeviceContext()
length = 1024
input_device = ctx.enqueue_create_buffer[DType.float32](https://mojolang.org/docs/manual/gpu/length.md)
# Initialize the input
with input_device.map_to_host() as input_host:
for i in range(length):
input_host[i] = Float32(i)
```
However, you should not use this in most production code because of the
bidirectional copies and synchronization. The example above is equivalent to:
```mojo
ctx = DeviceContext()
length = 1024
input_device = ctx.enqueue_create_buffer[DType.float32](https://mojolang.org/docs/manual/gpu/length.md)
input_host = ctx.enqueue_create_host_buffer[DType.float32](https://mojolang.org/docs/manual/gpu/length.md)
input_device.enqueue_copy_to(input_host)
ctx.synchronize()
for i in range(length):
input_host[i] = Float32(i)
input_host.enqueue_copy_to(input_device)
ctx.synchronize()
```
#### Deallocating memory buffers
Both `DeviceBuffer` and `HostBuffer` are subject to Mojo's standard ownership
and lifecycle mechanisms. The Mojo compiler analyzes our program to determine
the last point that the owner of or a reference to an object is used and
automatically adds a call to the object's destructor. This means that you don't
explicitly call any method to free the memory represented by a `DeviceBuffer` or
`HostBuffer` instance. See the [Ownership](/docs/manual/values/ownership/) and
[Intro to value lifecycle](/docs/manual/lifecycle/) sections of the Mojo Manual
for more information on Mojo value ownership and value lifecycle management, and
the [Death of a value](/docs/manual/lifecycle/death/) section for a detailed
explanation of object destruction.
### Compiling and enqueuing a kernel function for execution
The
[`compile_function()`](/docs/std/gpu/host/device_context/DeviceContext/#compile_function)
method accepts a kernel function as a compile-time
[parameter](/docs/manual/parameters/) and then compiles it for the associated
`DeviceContext`. Then you can enqueue the compiled kernel for execution by
passing it to the
[`enqueue_function()`](/docs/std/gpu/host/device_context/DeviceContext/#enqueue_function)
method. The example in the [GPU programming model](#gpu-programming-model)
demonstrated this pattern:
```mojo title="scalar_add.mojo"
...
scalar_add_kernel = ctx.compile_function[scalar_add]()
ctx.enqueue_function(
scalar_add_kernel,
device_buffer,
num_elements,
Float32(20.0),
grid_dim=1,
block_dim=num_elements,
)
```
When using a compiled kernel function like this, you execute it by calling
`enqueue_function()` with the following arguments in this order:
- The kernel function to execute.
- Any additional arguments specified by the kernel function definition in the
order specified by the function.
- The grid dimensions using the `grid_dim` keyword argument.
- The thread block dimensions using the `block_dim` keyword argument.
Refer to the [Multidimensional grids and thread
organization](#multidimensional-grids-and-thread-organization) section for more
information on grid and thread block dimensions.
:::note Compile-time type checking
The `compile_function()` and `enqueue_function()` methods type-check the
kernel function arguments at compile time. If an argument's type doesn't match
the kernel function's parameter list, you'll get a compile-time error instead
of a run-time failure.
:::
The advantage of compiling the kernel as a separate step is that you can
execute the same compiled kernel on the same device multiple times. This avoids
the overhead of compiling the kernel each time it's executed.
If your application needs to execute a kernel function only once, you can use an
overloaded version of `enqueue_function()` that compiles the kernel and
enqueues it in a single step. Therefore, the following is equivalent to the
separate calls to `compile_function()` and `enqueue_function()`
shown above (note that the kernel function is provided as a compile-time
parameter in this case):
```mojo
ctx.enqueue_function[scalar_add](https://mojolang.org/docs/manual/gpu/device_buffer,
num_elements,
Float32(20.0),
grid_dim=1,
block_dim=num_elements,
)
```