blob: a06bbf68aee2bc796df5fade270174c81222524c [file] [log] [blame]
Name Strings
Geoff Stahl, Apple Computer (gstahl 'at'
$Date: 2009/01/07 22:05:21 $ $Revision: 1.13 $
APPLE_fence can affect this extension.
This extension is designed to allow very high vertex processing rates which
are facilitated both by relieving the CPU of as much processing burden as
possible and by allowing graphics hardware to directly access vertex data.
Because this extension is implemented as an addition to the vertex array
specification provided by OpenGL 1.1, applications can continue to use
existing vertex submission logic while taking advantage of vertex array
ranges to more efficiently process those arrays.
The vertex array coherency model provided by OpenGL 1.1 requires that
vertex data specified in vertex arrays be transferred from system memory
each time Begin, DrawArrays, or DrawElements is called. Further, OpenGL
1.1 requires that the transfer of data be completed by the time End,
DrawArrays, or DrawElements returns. Both of these requirements are
relaxed by the vertex array range extension. Vertex data may be cached
by the GL so there is no guarantee that changes to the vertex data will
be reflected in following drawing commands unless it is flushed with
FlushVertexArrayRangeAPPLE. The reading of vertex data may be deferred
by the GL so there is no guarantee that the GL will be finished reading
the data until completion is forced by the use of Finish or the APPLE_fence
Vertex array range can be enabled in two ways. EnableClientState can be
used with the VERTEX_ARRAY_RANGE_APPLE param to enable vertex array range
for the client context. One can also simply set the vertex array storage
hint to either STORAGE_CACHED_APPLE or STORAGE_SHARED_APPLE (as discussed
below) to enable a particular vertex array range. Once this is done, use of
vertex array range requires the definition of a specific memory range for
vertex data through VertexArrayRangeAPPLE. It is recommended this data be
page aligned (4096 byte boundaries) and a multiple of page size in length
for maximum efficiency in data handling and internal flushing, but this is
not a requirement and any location and length of data can be defined as a
vertex array. This extension provides no memory allocators as any
convenient memory allocator can be used.
Once a data set is established, using VertexArrayRangeAPPLE, it can be can
be drawn using standard OpenGL vertex array commands, as one would do
without this extension. Note, if any the data for any enabled array for a
given array element index falls outside of the vertex array range, an
undefined vertex is generated. One should also understand removing or
replacing all calls to vertex array range functions with no-ops or disabling
the vertex array range by disabling the VERTEX_ARRAY_RANGE_APPLE client
state should not change the results of an application's OpenGL drawing.
For static data no additional coherency nor synchronization must be done and
the client is free to draw with the specified draw as it sees fit.
If data is dynamic, thus to be modified, FlushVertexArrayRangeAPPLE should
be used. The command is issued when data has been modified since the last
call to VertexArrayRangeAPPLE or FlushVertexArrayRangeAPPLE and prior to
drawing with such data. FlushVertexArrayRangeAPPLE only provides memory
coherency prior to drawing (such as ensuring CPU caches are flushed or VRAM
cached copies are updated) and does not provide any synchronization with
previously issued drawing commands. The range flushed can be the specific
range modified and does not have to be the entire vertex array range.
Additionally, data maybe read immediately after a flush without need for
further synchronization, thus overlapping areas of data maybe read, modified
and written between two successive flushes and the data will be
To synchronize data modification after drawing two methods can be used. A
Finish command can be issued which will not return until all previously
issued commands are complete, forcing completely synchronous operation.
While this guarantees all drawing is complete it may not be the optimal
solution for clients which just need to ensure drawing with the vertex array
range or a specific range with the array is compete. The APPLE_fence
extension can be used when dynamic data modifications need to be
synchronized with drawing commands. Specifically, if data is to be modified,
a fence can be set immediately after drawing with the data. Once it comes
time to modify the data, the application must test (or finish) this fence to
ensure the drawing command has completed. Failure to do this could result in
new data being used by the previously issued drawing commands. It should be
noted that providing the maximum time between the drawing set fence and the
modification test/finish fence allows the most asynchronous behavior and
will result in the least stalling waiting for drawing completion. Techniques
such as double buffering vertex data can be used to help further prevent
stalls based on fence completion but are beyond the scope of this extension.
Once an application is finished with a specific vertex array range or at
latest prior to exit, and prior to freeing the memory associated with this
vertex array, the client should call VertexArrayRangeAPPLE with a data
location and length of 0 to allow the internal memory managers to complete
any commitments for the array range. In this case once
VertexArrayRangeAPPLE returns it is safe to de-allocate the memory.
Three types of storage hints are available for vertex array ranges; client,
shared, and cached. These hints are set by passing the
VertexArrayParameteriAPPLE with VERTEX_ARRAY_STORAGE_HINT_APPLE pname.
Client storage, the default OpenGL behavior, occurs when
set. Note, STORAGE_CLIENT_APPLE is also the default hint setting. Shared
memory usage is normally used for dynamic data that is expected to be
modified and is likely mapped to AGP memory space for access by both the
graphics hardware and client. It is set when either
being set, or in all cases when the STORAGE_SHARED_APPLE hint is set.
Finally, the cached storage is designed to support static data and data which
could be cached in VRAM. This provides maximum access bandwidth for the
vertex array and occurs when the STORAGE_CACHED_APPLE hint is set.
The following pseudo-code represents the treatment of a vertex array range
memory depending on the hint setting and whether vertex array range is
enabled for the client context:
vertex array is treated as cached
vertex array is treated as shared
else if (VERTEX_ARRAY_RANGE_APPLE enabled)
vertex array is treated as shared
vertex array is treated as client
Note, these hints can affect how array flushes are handled and the overhead
associated with flushing the array, it is recommended that data be handled
as shared unless it really is static and there are no plans to modify it.
To summarize the vertex array range extension provides relaxed
synchronization rules for handling vertex array data allowing high bandwidth
asynchronous data transfer from client memory to graphics hardware.
Different flushing and synchronization rules are required to ensure data
coherency when modifying data. Lastly, memory handling hints are provided
to allow the tunning of memory storage and access for maximum efficiency.
How does one get the current VERTEX_ARRAY_STORAGE_HINT_APPLE (storage hint)
for a vertex array range?
via GetIntegerv with VERTEX_ARRAY_STORAGE_HINT_APPLE as the pname.
How does this extension interact with the compiled_vertex_array extension?
RESOLUTION: They are independent and not interfere with each other. In
practice, if you use APPLE_vertex_array_range, you can surpass the
performance of compiled_vertex_array
Should we give a programmer a sense of how big a vertex array range they can
RESOLUTION: Not completely resolved. There should be a query for
determining the maximum safe size of a vertex array range which, we do
not have as of yet. Currently, the vertex array range size plus the
command buffers must not exceed available GART space, which in real world
terms, means the smallest maximum vertex array range size is about 24 MB
on any currently supported hardware on Mac OS X and get's larger normally
with more RAM and/or more advanced graphics hardware. Failure modes are
likely either failure to speed up the vertex processing or failure to
draw all the vertex data. So clients should plan on single vertex array
ranges of less than 24 MB if they intend to run on all Mac OS X hardware.
Should Flush be the same as FlushVertexArrayRangeAPPLE?
RESOLUTION: No. A Flush is a different concept than
FlushVertexArrayRangeAPPLE. a Flush submits pending OpenGL commands to the
OpenGL engine for processing while, a FlushVertexArrayRangeAPPLE just ensures
memory coherency for the vertex array range and does not perform a Flush
in OpenGL terms.
New Procedures and Functions
void VertexArrayRangeAPPLE(sizei length, void *pointer);
void FlushVertexArrayRangeAPPLE(sizei length, void *pointer);
void VertexArrayParameteriAPPLE(enum pname, int param);
New Tokens
Accepted by the <cap> parameter of EnableClientState, DisableClientState,
and IsEnabled:
Accepted by the <pname> parameter of GetBooleanv, GetIntegerv, GetFloatv,
and GetDoublev:
Accepted by the <pname> parameter of GetPointerv:
Accepted by the <pname> parameter of VertexArrayParameteriAPPLE,
GetBooleanv, GetIntegerv, GetFloatv, and GetDoublev:
Accepted by the <param> parameter of VertexArrayParameteriAPPLE:
Additions to Chapter 2 of the OpenGL 1.1 Specification (OpenGL Operation)
After the discussion of vertex arrays (Section 2.8) add a
description of the vertex array range:
"The command
void VertexArrayRangeAPPLE(sizei length, void *pointer)
specifies the current vertex array range. When the vertex array range is
specified and valid, vertex transfers from within the vertex array
range are potentially faster. The vertex array range is a contiguous
region of address space for placing vertex arrays. The "pointer" parameter
is a pointer to the base of the vertex array range. The "length" pointer is
the length of the vertex array range in basic machine units (typically
unsigned bytes). Memory associated with a vertex array range should be
allocated by the client and the responsibility for maintaining it rests with
the client as long as it is being used as a vertex array range.
The vertex array range address space region extends from "pointer" to
"pointer + length - 1" inclusive. When specified, vertex array vertex
transfers from within the vertex array range are potentially faster.
There is some system burden associated with establishing a vertex array
range (typically, the memory range must be locked down). If either the
vertex array range pointer or size is set to zero, the previously
established vertex array range is released (typically, unlocking the
memory). This should always be done prior to freeing of the memory by the
The vertex array range may not be established for operating system dependent
reasons, and therefore, not valid. Reasons that a vertex array range cannot
be established include exceeding the maximum vertex array range size, the
memory could not be locked down, etc.
The vertex array range is considered enabled after VERTEX_ARRAY_RANGE_APPLE
client state is enabled and as soon as a valid vertex array range is
specified and disabled once the size length and/or pointer is set to zero or
VERTEX_ARRAY_RANGE_APPLE client state is disabled.
When the vertex array range is enabled, ArrayElement commands may generate
undefined vertices if and only if any indexed elements of the enabled arrays
are not within the vertex array range or if the index is negative or greater
or equal to the implementation-dependent value of
MAX_VERTEX_ARRAY_RANGE_ELEMENT_APPLE. If an undefined vertex is generated,
an INVALID_OPERATION error may or may not be generated.
The vertex array coherency model specifies when vertex data must be extracted
from the vertex array memory. When the vertex array range is not valid,
(quoting the specification) `Changes made to array data between the
execution of Begin and the corresponding execution of End may effect calls
to ArrayElement that are made within the same Begin/End period in
non-sequential ways. That is, a call to ArrayElement that precedes a change
to array data may access the changed data, and a call that follows a change
to array data may access the original data.'
When the vertex array range is valid, the vertex array coherency model is
relaxed so that changes made to array data may affect calls to ArrayElement
in non-sequential ways. That is a call to ArrayElement that precedes a
change to array data may access the changed data, and a call that follows a
change to array data may access original data. This requires in two points
of synchronization to maintain coherency.
The first point where synchronization must occur to maintain coherency is
post data modification, prior to drawing. FlushVertexArrayRangeAPPLE should
be used by the client on all ranges of memory which have been modified since
the last call to VertexArrayRangeAPPLE or FlushVertexArrayRangeAPPLE.
The second point of synchronization is after drawing with a vertex array
range and prior to modifying it's data. In this case either Finish or a
fence must be used. Finish will create a synchronization point for all
drawing an may not be the optimal method to ensure drawing completion prior
to data modification. A fence, defined in the APPLE_fence extension, on the
other hand allows more selective synchronization. The client can set a fence
immediately after drawing with the data in question and test or finish that
fence prior to modifying the data. See the APPLE_fence extension for more
To maintain full coherency, once a vertex array range is enabled, requires
the client to both flush the vertex array after data modification, prior to
drawing, and synchronize with Finish or a fence after drawing, prior to
modifying the data.
The command
void VertexArrayParameteriAPPLE(enum pname, int param)
allows the client to hint at the expected use of the vertex array range.
pname must be VERTEX_ARRAY_STORAGE_HINT_APPLE. param can either be
STORAGE_CACHED_APPLE or STORAGE_SHARED_APPLE and can be used by the system
to tune the handling of the vertex array range data. These
parameters are just hints and require no specific handling by the system.
The default state is STORAGE_SHARED_APPLE which, indicates that the vertex
data is expected to be dynamic and should be handled in a way to optimize
modification and flushing of the vertex array range, if possible.
STORAGE_CACHED_APPLE indicates the data is expected to static and techniques
such as VRAM caching could be employed to optimize memory bandwidth to the
vertex array range. Proper use of FlushVertexArrayRangeAPPLE guarantees
memory coherency in all cases and will result in deterministic defined
behavior in all cases, whether hints are employed or not.
The client state required to implement the vertex array range consists of an
enable bit, a memory pointer, an integer size, and a valid bit."
Additions to Chapter 5 of the OpenGL 1.1 Specification (Special Functions)
Add to the end of Section 5.4 "Display Lists"
"VertexArrayRangeAPPLE and FlushVertexArrayRangeAPPLE are not complied into
display lists but are executed immediately.
If a display list is compiled while VERTEX_ARRAY_RANGE_APPLE is enabled, the
commands ArrayElement, DrawArrays, DrawElements, and DrawRangeElements are
accumulated into a display list as if VERTEX_ARRAY_RANGE_APPLE is disabled."
Additions to the CGL interface:
Additions to the WGL interface:
Additions to the GLX Specification
GLX Protocol
INVALID_OPERATION is generated if VertexArrayRangeAPPLE or
FlushVertexArrayRangeAPPLE is called between the execution of Begin and the
corresponding execution of End.
INVALID_OPERATION may be generated if an undefined vertex is generated.
New State
Get Value Get Command Type Value Attrib
--------- ----------- ---- ------- ------------
VERTEX_ARRAY_RANGE_APPLE IsEnabled B False vertex-array
VERTEX_ARRAY_RANGE_POINTER_APPLE GetPointerv Z+ 0 vertex-array
VERTEX_ARRAY_RANGE_LENGTH_APPLE GetIntegerv Z+ 0 vertex-array
New Implementation Dependent State
Get Value Get Command Type Minimum Value
--------- ----------- ----- -------------
Revision History