File buffer.h

This file contains the interface definition for the backends.

For normal use you should not call the functions defined in this file directly.

See
array.h For managing buffers
See
kernel.h For using kernels

Defines

GA_CTX_SCHED_AUTO

Automatic scheduling, decide what to do depending on the workload, number of cores in the computer and other relevant factors. (default)

GA_CTX_SCHED_SINGLE

Single-work scheduling. Optimize for speed in a single process, with a single thread. This is the fastest mode, but it may keep the CPU busy more than necessary.

GA_CTX_SCHED_MULTI

Multi-work scheduling. Try to not keep the CPU busy more than necessary and let other threads a chance at some CPU time. This may increase the latency when waiting for GPU operations.

GA_BUFFER_READ_WRITE

The buffer is available for reading and writing from kernels.

This is the default (0) value.

GA_BUFFER_DEV

Allocate the buffer in device-only memory.

This is the default (0) value.

GA_BUFFER_READ_ONLY

Signal that the memory in this buffer will only be read by kernels.

You can use gpudata_write() to set the contents.

You may not call gpudata_memset() with the resulting buffer as the destination.

GA_BUFFER_WRITE_ONLY

Signal that the memory in this buffer will only be written by kernels (i.e. it is an output buffer).

You can read the contents with gpudata_read().

GA_BUFFER_INIT

Initialize the contents of the buffer with the user-supplied host buffer (data). This buffer must be at least sz large.

GA_BUFFER_HOST

Allocate the buffer in host-reachable memory enabling you to retrieve a pointer to the contents as the GA_BUFFER_PROP_HOSTPOINTER property.

GA_BUFFER_MASK
GA_CTX_PROP_DEVNAME

Get the device name for the context.

Type: char [256]

GA_CTX_PROP_LMEMSIZE

Get the local memory size available for a call in the context.

Type: size_t

GA_CTX_PROP_NUMPROCS

Number of compute units in this context.

compute units times local size is more or less the expected parallelism available on the device, but this is a very rough estimate.

Type: unsigned int

GA_CTX_PROP_BIN_ID

Get the compatibility ID for the binaries generated with this context.

Those binaries should work with any context which has the same ID.

Type: const char *

GA_CTX_PROP_ERRBUF

Get a pre-allocated 8 byte buffer for kernel ops.

This buffer is initialized to 0 on allocation and must always be returned to that state after using it.

This only to avoid the overhead of an allocation when calling a kernel that may error out. It does not preclude the need for synchronization and transfers.

Type: gpudata *

GA_CTX_PROP_TOTAL_GMEM

Get the total size of global memory on the device.

Type: size_t

GA_CTX_PROP_FREE_GMEM

Get the size of free global memory on the device.

Type: size_t

GA_CTX_PROP_NATIVE_FLOAT16

Get the status of native float16 support on the device.

Type: int

GA_CTX_PROP_MAXGSIZE0

Get the maximum global size for dimension 0.

Type: size_t

GA_CTX_PROP_MAXGSIZE1

Get the maximum global size for dimension 1.

Type: size_t

GA_CTX_PROP_MAXGSIZE2

Get the maximum global size for dimension 2.

Type: size_t

GA_CTX_PROP_MAXLSIZE0

Get the maximum local size for dimension 0.

Type: size_t

GA_CTX_PROP_MAXLSIZE1

Get the maximum local size for dimension 1.

Type: size_t

GA_CTX_PROP_MAXLSIZE2

Get the maximum loca size for dimension 2.

Type: size_t

GA_CTX_PROP_UNIQUE_ID

Get a unique ID for the device behind the context.

Type: char [16]

GA_CTX_PROP_LARGEST_MEMBLOCK

Get the largest single block of memory that can be allocted.

Type: size_t

GA_BUFFER_PROP_START
GA_BUFFER_PROP_CTX

Get the context in which this buffer was allocated.

Type: gpucontext *

GA_BUFFER_PROP_REFCNT

The reference count of the buffer. Use only for debugging purposes.

Type: unsigned int

GA_BUFFER_PROP_SIZE

Size of the buffer on the device.

This may be larger than the requested allocation size due to a number of factors.

Type: size_t

GA_KERNEL_PROP_START
GA_KERNEL_PROP_CTX

Get the context for which this kernel was compiled.

Type: gpucontext *

GA_KERNEL_PROP_MAXLSIZE

Get the maximum block size (also known as local size) for a call of this kernel.

Type: size_t

GA_KERNEL_PROP_PREFLSIZE

Get the prefered multiple of the block size for a call to this kernel.

Type: size_t

GA_KERNEL_PROP_NUMARGS

Get the number of kernel arguments.

Type unsigned int

GA_KERNEL_PROP_TYPES

Get the list of argument types for a kernel.

This list is the same length as the number of arguments to the kernel. Do not modify the returned list.

Type: const int *

Typedefs

typedef struct _gpucontext gpucontext

Opaque struct for context data.

typedef struct _gpucontext_props gpucontext_props

Opaque structure that holds properties for the context.

Enums

_ga_usefl

Flags for gpukernel_init().

It is important to specify these properly as the compilation machinery will ensure that the proper configuration is made to support the requested features or error out if the demands cannot be met.

Warning
Failure to properly specify the feature flags will in most cases result in silent data corruption (especially on ATI cards).

Values:

0x02

The kernel makes use of small (size is smaller than 4 bytes) types.

0x04

The kernel makes use of double or complex doubles.

0x08

The kernel makes use of complex of complex doubles.

0x10

The kernel makes use of half-floats (also known as float16)

0x2000

The kernel is made of CUDA code.

0x4000

The kernel is made of OpenCL code.

Functions

int gpu_get_platform_count(const char * name, unsigned int * platcount)

Gets information about the number of available platforms for the backend specified in name.

Return
GA_NO_ERROR, if success
Parameters
  • name: the backend name
  • platcount: will contain number of compatible platforms in host

int gpu_get_device_count(const char * name, unsigned int platform, unsigned int * devcount)

Gets information about the number of compatible devices on a specific host’s platform for the backend specified in name.

Return
GA_NO_ERROR, if success
Parameters
  • name: the backend name
  • platform: number for a platform in host
  • devcount: will contain number of compatible devices in platform

int gpucontext_props_new(gpucontext_props ** res)

Allocate and initialized an instance of gpucontext_props.

Initialization is done with default values.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • res: pointer to storage space for the created object

int gpucontext_props_cuda_dev(gpucontext_props * p, int devno)

Set the device number for a CUDA device.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • p: properties object
  • devno: device number

int gpucontext_props_opencl_dev(gpucontext_props * p, int platno, int devno)

Set the platform and device for OpenCL.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • p: properties object
  • platno: platform number
  • devno: device number

int gpucontext_props_sched(gpucontext_props * p, int sched)

Set the scheduling mode for the device.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • p: properties object
  • sched: scheduling mode. One of these.

int gpucontext_props_set_single_stream(gpucontext_props * p)

Set single-stream mode.

All operations on the device will be serialized on a single stream. This will also disable most of the interlocking normally done between multiple streams to keep everything in order.

This mode can be faster if you don’t have a lot of device-level parallelism in your workload.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • p: properties object

int gpucontext_props_kernel_cache(gpucontext_props * p, const char * path)

Set the path for the kernel cache.

The cache can be shared with other running instances, even on shared drives.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • p: properties object
  • path: desired location of the kernel cache

int gpucontext_props_alloc_cache(gpucontext_props * p, size_t initial, size_t max)

Configure the allocation cache.

The maximum size is also a limit on the total amount of memory allocated on the device.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • p: properties object
  • initial: initial size of the cache
  • max: maximum size of the cache

void gpucontext_props_del(gpucontext_props * p)

Free a properties object.

This should not be called on a properties object that has been passed to gpucontext_init().

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • p: properties object

int gpucontext_init(gpucontext ** res, const char * name, gpucontext_props * props)

Create a context on the specified device.

The passed-in properties pointer will be managed by this function and needs not be freed. This means that you shouldn’t touch the properties object after passing it to this function.

Warning
This function is not thread-safe.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • res: a pointer to a location that will be allocated
  • name: the backend name.
  • props: a properties object for the context. Can be NULL for defaults.

void gpucontext_deref(gpucontext * ctx)

Dereference a context.

This removes a reference to the context and as soon as the reference count drops to zero the context is destroyed. The context can stay alive after you call this function because some object keep a reference to their context.

Parameters
  • ctx: a valid context pointer.

int gpucontext_property(gpucontext * ctx, int prop_id, void * res)

Fetch a context property.

The property must be a context property. The currently defined properties and their type are defined in Properties.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • ctx: context
  • prop_id: property id (from Properties)
  • res: pointer to the return space of the appropriate type

const char* gpucontext_error(gpucontext * ctx, int err)

Get a string describing err.

If you need to get a description of a error that occurred during context creation, call this function using NULL as the context. This version of the call is not thread-safe.

Return
string description of error
Parameters
  • ctx: the context in which the error occured
  • err: error code

gpudata* gpudata_alloc(gpucontext * ctx, size_t sz, void * data, int flags, int * ret)

Allocates a buffer of size sz in context ctx.

Buffers are reference counted internally and start with a reference count of 1.

Return
A non-NULL pointer to a gpudata structure. This structure is intentionally opaque as its content may change according to the backend used.
Parameters
  • ctx: a context pointer
  • sz: the requested size
  • flags: see Allocation flags
  • data: optional pointer to host buffer
  • ret: error return pointer

void gpudata_retain(gpudata * b)

Increase the reference count to the passed buffer by 1.

Parameters
  • b: a buffer

void gpudata_release(gpudata * b)

Release a buffer.

This will decrement the reference count of the buffer by 1. If that count reaches 0 all associated ressources will be released.

Even if your application does not have any references left to a buffer it may still hang around if it is in use by internal mechanisms (kernel call, …)

int gpudata_share(gpudata * a, gpudata * b, int * ret)

Check if two buffers may overlap.

Both buffers must have been created with the same backend.

Parameters
  • a: first buffer
  • b: second buffer
  • ret: error return pointer
Return Value
  • 1: The buffers may overlap
  • 0: The buffers do not overlap.
  • -1: An error was encoutered, ret contains a detailed error code if not NULL.

int gpudata_move(gpudata * dst, size_t dstoff, gpudata * src, size_t srcoff, size_t sz)

Copy the content of a buffer to another.

Both buffers must be in the same context and contiguous. Additionally the buffers must not overlap otherwise the content of the destination buffer is not defined.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • dst: destination buffer
  • dstoff: offset inside the destination buffer
  • src: source buffer
  • srcoff: offset inside the source buffer
  • sz: size of data to copy (in bytes)

int gpudata_transfer(gpudata * dst, size_t dstoff, gpudata * src, size_t srcoff, size_t sz)

Transfer the content of buffer across contexts.

If possible it will try to the the transfer in an efficient way using backend-specific tricks. If those fail or can’t be used, it will fallback to a copy through the host.

Return
the new buffer in dst_ctx or NULL if no efficient way to transfer could be found.
Parameters
  • dst: buffer to transfer to
  • dstoff: offset in the destination buffer
  • src: buffer to transfer from
  • srcoff: offset in the source buffer
  • sz: size of the region to transfer

int gpudata_read(void * dst, gpudata * src, size_t srcoff, size_t sz)

Transfer data from a buffer to memory.

The buffer and the memory region must be contiguous.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • dst: destination in memory
  • src: source buffer
  • srcoff: offset inside the source buffer
  • sz: size of data to copy (in bytes)

int gpudata_write(gpudata * dst, size_t dstoff, const void * src, size_t sz)

Transfer data from memory to a buffer.

The buffer and the memory region must be contiguous.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • dst: destination buffer
  • dstoff: offset inside the destination buffer
  • src: source in memory
  • sz: size of data to copy (in bytes)

int gpudata_memset(gpudata * dst, size_t dstoff, int data)

Set a buffer to a byte pattern.

This function acts like the C function memset() for device buffers.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • dst: destination buffer
  • dstoff: offset into the destination buffer
  • data: byte value to write into the destination.

int gpudata_sync(gpudata * b)

Synchronize a buffer.

Waits for all previous read, writes, copies and kernel calls involving this buffer to be finished.

This call is not required for normal use of the library as all exposed operations will properly synchronize amongst themselves. This call may be useful in a performance timing context to ensure that the work is really done, or before interaction with another library to wait for pending operations.

int gpudata_property(gpudata * buf, int prop_id, void * res)

Fetch a buffer property.

Can be used for buffer properties and context properties. Context properties will fetch the value for the context associated with the buffer. The currently defined properties and their type are defined in Properties.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • buf: buffer
  • prop_id: property id (from Properties)
  • res: pointer to the return space of the appropriate type

gpucontext* gpudata_context(gpudata * b)
gpukernel* gpukernel_init(gpucontext * ctx, unsigned int count, const char ** strings, const size_t * lengths, const char * fname, unsigned int numargs, const int * typecodes, int flags, int * ret, char ** err_str)

Compile a kernel.

Compile the kernel composed of the concatenated strings in strings and return a callable kernel. If lengths is NULL then all the strings must be NUL-terminated. Otherwise, it doesn’t matter (but the lengths must not include the final NUL byte if provided).

If

*err_str is not NULL on return, the caller must call free(*err_str) after use.
Parameters
  • ctx: context to work in
  • count: number of input strings
  • strings: table of string pointers
  • lengths: (optional) length for each string in the table
  • fname: name of the kernel function (as defined in the code)
  • numargs: number of kernel arguments
  • typecodes: the type of each argument
  • flags: flags for compilation (see #ga_usefl)
  • ret: error return pointer
  • err_str: returns pointer to debug message from GPU backend (if provided a non-NULL err_str)

Return
Allocated kernel structure or NULL if an error occured. ret will be updated with the error code if not NULL.

void gpukernel_retain(gpukernel * k)

Retain a kernel.

Increase the reference count of the passed kernel by 1.

Parameters
  • k: a kernel

void gpukernel_release(gpukernel * k)

Release a kernel.

Decrease the reference count of a kernel. If it reaches 0, all resources associated with k will be released.

If the reference count of a kernel reaches 0 while it is running, this call will block until completion.

int gpukernel_setarg(gpukernel * k, unsigned int i, void * a)

Set kernel argument.

Buffer arguments will not be retained and it is the responsability of the caller to ensure that the value is still valid whenever a call is made.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • k: kernel
  • i: argument index (starting at 0)
  • a: pointer to argument

int gpukernel_call(gpukernel * k, unsigned int n, const size_t * gs, const size_t * ls, size_t shared, void ** args)

Call a kernel.

If args is NULL, it will be assumed that the arguments have previously been set with kernel_setarg().

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • k: kernel
  • n: number of dimensions of grid/block
  • gs: grid sizes for this call (also known as global size)
  • ls: block sizes for this call (also known as local size)
  • shared: amount of dynamic shared memory to reserve
  • args: table of pointers to each argument (optional).

int gpukernel_property(gpukernel * k, int prop_id, void * res)

Fetch a property.

Can be used for kernel and context properties. The context properties will fetch the value for the context associated with the kernel. The currently defined properties and their type are defined in Properties.

Return
GA_NO_ERROR or an error code if an error occurred.
Parameters
  • k: kernel
  • prop_id: property id (from Properties)
  • res: pointer to the return space of the appropriate type

gpucontext* gpukernel_context(gpukernel * k)