4.2. CUDA Kernel API

4.2.1. Kernel declaration

The @cuda.jit decorator is used to create a CUDA kernel:

numba.cuda.jit(func_or_sig=None, argtypes=None, device=False, inline=False, bind=True, link=[], debug=None, **kws)

JIT compile a python function conforming to the CUDA Python specification. If a signature is supplied, then a function is returned that takes a function to compile. If

Parameters:
  • func_or_sig (function or numba.typing.Signature) –

    A function to JIT compile, or a signature of a function to compile. If a function is supplied, then an AutoJitCUDAKernel is returned. If a signature is supplied, then a function which takes a function to compile and returns an AutoJitCUDAKernel is returned.

    Note

    A kernel cannot have any return value.

  • device (bool) – Indicates whether this is a device function.
  • bind (bool) – Force binding to CUDA context immediately
  • link (list) – A list of files containing PTX source to link with the function
  • debug – If True, check for exceptions thrown when executing the kernel. Since this degrades performance, this should only be used for debugging purposes. Defaults to False. (The default value can be overriden by setting environment variable NUMBA_CUDA_DEBUGINFO=1.)
  • fastmath – If true, enables flush-to-zero and fused-multiply-add, disables precise division and square root. This parameter has no effect on device function, whose fastmath setting depends on the kernel function from which they are called.
class numba.cuda.compiler.AutoJitCUDAKernel(func, bind, targetoptions)

CUDA Kernel object. When called, the kernel object will specialize itself for the given arguments (if no suitable specialized version already exists) and launch on the device associated with the current context.

Kernel objects are not to be constructed by the user, but instead are created using the numba.cuda.jit() decorator.

inspect_asm(signature=None)

Return the generated assembly code for all signatures encountered thus far, or the LLVM IR for a specific signature if given.

inspect_llvm(signature=None)

Return the LLVM IR for all signatures encountered thus far, or the LLVM IR for a specific signature if given.

inspect_types(file=None)

Produce a dump of the Python source of this function annotated with the corresponding Numba IR and type information. The dump is written to file, or sys.stdout if file is None.

specialize(*args)

Compile and bind to the current context a version of this kernel specialized for the given args.

Individual specialized kernels are instances of numba.cuda.compiler.CUDAKernel:

class numba.cuda.compiler.CUDAKernel(llvm_module, name, pretty_name, argtypes, call_helper, link=(), debug=False, fastmath=False, type_annotation=None)

CUDA Kernel specialized for a given set of argument types. When called, this object will validate that the argument types match those for which it is specialized, and then launch the kernel on the device.

bind()

Force binding to current CUDA context

device

Get current active context

inspect_asm()

Returns the PTX code for this kernel.

inspect_llvm()

Returns the LLVM IR for this kernel.

inspect_types(file=None)

Produce a dump of the Python source of this function annotated with the corresponding Numba IR and type information. The dump is written to file, or sys.stdout if file is None.

ptx

PTX code for this kernel.

4.2.2. Intrinsic Attributes and Functions

The remainder of the attributes and functions in this section may only be called from within a CUDA Kernel.

4.2.2.1. Thread Indexing

numba.cuda.threadIdx

The thread indices in the current thread block, accessed through the attributes x, y, and z. Each index is an integer spanning the range from 0 inclusive to the corresponding value of the attribute in numba.cuda.blockDim exclusive.

numba.cuda.blockIdx

The block indices in the grid of thread blocks, accessed through the attributes x, y, and z. Each index is an integer spanning the range from 0 inclusive to the corresponding value of the attribute in numba.cuda.gridDim exclusive.

numba.cuda.blockDim

The shape of a block of threads, as declared when instantiating the kernel. This value is the same for all threads in a given kernel, even if they belong to different blocks (i.e. each block is “full”).

numba.cuda.gridDim

The shape of the grid of blocks, accressed through the attributes x, y, and z.

numba.cuda.grid(ndim)

Return the absolute position of the current thread in the entire grid of blocks. ndim should correspond to the number of dimensions declared when instantiating the kernel. If ndim is 1, a single integer is returned. If ndim is 2 or 3, a tuple of the given number of integers is returned.

Computation of the first integer is as follows:

cuda.threadIdx.x + cuda.blockIdx.x * cuda.blockDim.x

and is similar for the other two indices, but using the y and z attributes.

numba.cuda.gridsize(ndim)

Return the absolute size (or shape) in threads of the entire grid of blocks. ndim should correspond to the number of dimensions declared when instantiating the kernel.

Computation of the first integer is as follows:

cuda.blockDim.x * cuda.gridDim.x

and is similar for the other two indices, but using the y and z attributes.

4.2.2.2. Memory Management

numba.cuda.shared.array(shape, dtype)

Creates an array in the local memory space of the CUDA kernel with the given shape and dtype.

Returns an array with its content uninitialized.

Note

All threads in the same thread block sees the same array.

numba.cuda.local.array(shape, dtype)

Creates an array in the local memory space of the CUDA kernel with the given shape and dtype.

Returns an array with its content uninitialized.

Note

Each thread sees a unique array.

numba.cuda.const.array_like(ary)

Copies the ary into constant memory space on the CUDA kernel at compile time.

Returns an array like the ary argument.

Note

All threads and blocks see the same array.

4.2.2.3. Synchronization and Atomic Operations

numba.cuda.atomic.add(array, idx, value)

Perform array[idx] += value. Support int32, int64, float32 and float64 only. The idx argument can be an integer or a tuple of integer indices for indexing into multiple dimensional arrays. The number of element in idx must match the number of dimension of array.

Returns the value of array[idx] before the storing the new value. Behaves like an atomic load.

numba.cuda.atomic.max(array, idx, value)

Perform array[idx] = max(array[idx], value). Support int32, int64, float32 and float64 only. The idx argument can be an integer or a tuple of integer indices for indexing into multiple dimensional arrays. The number of element in idx must match the number of dimension of array.

Returns the value of array[idx] before the storing the new value. Behaves like an atomic load.

numba.cuda.syncthreads()

Synchronize all threads in the same thread block. This function implements the same pattern as barriers in traditional multi-threaded programming: this function waits until all threads in the block call it, at which point it returns control to all its callers.

Warning

Must be called by every thread in the thread-block. Falling to do so may result in undefined behavior.

4.2.2.4. Memory Fences

The memory fences are used to guarantee the effect of memory operations are visible by other threads within the same thread-block, the same GPU device, and the same system (across GPUs on global memory). Memory loads and stores are guaranteed to not move across the memory fences by optimization passes.

Warning

The memory fences are considered to be advanced API and most usercases should use the thread barrier (e.g. syncthreads()).

numba.cuda.threadfence()

A memory fence at device level (within the GPU).

numba.cuda.threadfence_block()

A memory fence at thread block level.

numba.cuda.threadfence_system()

A memory fence at system level (across GPUs)