This commit updates a *lot* of rather questionable path handling
logic to be robust. Specifically:
* All path operations go through Platform::Path.
* All ad-hoc path handling functions are removed, together with
PATH_SEP. This removes code that was in platform-independent
parts, but had platform-dependent behavior.
* Group::linkFileRel is removed; only an absolute path is stored
in Group::linkFile. However, only Group::linkFileRel is saved,
with the relative path calculated on the fly, from the filename
passed into SaveToFile. This eliminates dependence on global
state, and makes it unnecessary to have separare code paths
for saved and not yet saved files.
* In a departure from previous practice, functions with
platform-independent code but platform-dependent behavior
are all grouped under platform/. This makes it easy to grep
for functions with platform-dependent behavior.
* Similarly, new (GUI-independent) code for all platforms is added
in the same platform.cpp file, guarded with #ifs. It turns out
that implementations for different platforms had a lot of shared
code that tended to go out of sync.
The existing code is horrible and needlessly platform-dependent.
Even worse, it causes a freeze on GTK. Instead of propping that up
with a few more crutches, just fix the root cause.
Before this commit, the first time NewFrame() is called,
the background color would not be filled, leading to interference
with whatever the GUI toolkit decided to put there.
Before this commit, updates to the bitmap font in the graphics window
cause missing characters in the text window and vice versa. This is
because BitmapFont contains a texture that's mutated, and sharing it
would also require sharing display lists between GL contexts, which
is not done (and overly complicated for this case anyway).
Before this commit, dragging a new rect would result in one
of the lines lagging behind, because it is drawn in the middle
of regeneration. After this commit, the rectangle stays rectangular.
For a reason I do not understand, this only fixes Win32; GTK
continues to exhibit the bug, whereas Cocoa has never exhibited it
but the render latency seems to have lowered slightly.
Before this commit, when an entity is clicked at or dragged, and it
shares a place with other entities, which of them is selected is
decided more or less at random. This is particularly annoying when
dragging.
After this commit, when clicking, an entity from the current group
is given preference, and when dragging, an entity from a request
is given preference. This allows e.g. dragging points of a sketch
even when an extrusion of that sketch is active.
This commit does three things:
* Recognizes that BeginFrame()/EndFrame() are badly named, since
BeginFrame() sets up framebuffer, and EndFrame() flushes a frame,
and they do not have to be called in pairs; and so renames them
to NewFrame()/FlushFrame().
* Reduces the amount of frame flushes in GraphicsWindow::Paint()
to two, which is the minimum since we use two different cameras
for geometry and UI;
* Changes the FPS measurement code to only take into account
the time spent rendering our main geometry, and not the UI
rendering or window system interaction time.
To actually achieve improved performance with the OpenGL 2 renderer,
we have to cache geometry that doesn't change when the viewport does
(note that the rendered pixels can change quite dramatically because
we can reconfigure shaders; e.g. stippling can be drawn in screen
coordinates).
This commit adds a BatchCanvas interface that can be implemented
by renderers, and uses it for drawing entities such as lines and
points.
There are two main reasons to desire an OpenGL 2 renderer:
1. Compatibility. The compatibility profile, ironically, does not
offer a lot of compatibility, and our OpenGL 1 renderer will not
run on Android, iOS, or WebGL.
2. Performance. The immediate mode does not scale, and in fact
becomes very slow with only a moderate amount of lines on screen,
and only a somewhat large amount of triangles.
This commit implements a basic OpenGL 2 renderer that uses only
features from the (OpenGL 3.2) core profile. It is not yet faster
than the OpenGL 1 renderer, primarily because it uses a lot of small
draw calls.
This commit uses OpenGL 2 on Linux and Mac OS X directly (i.e. links
to the GL symbols from version 2+); on Windows this is impossible
with the default drivers, so for now OpenGL 1 is still used there.
Abstract the exact details of the OpenGL renderer in the render.h
header; this allows us to use GL-specific types in the renderer
class and functions without including OpenGL (and Windows, where
applicable) headers in every source file.
The states are:
* Draw all lines (on top of shaded mesh).
* Draw occluded (by shaded mesh) lines as stippled.
* Do not draw occluded (by shaded mesh) lines.
As usual, the export output follows the screen output.
This commit alters the build system substantially; it adds another
platform, `headless`, that provides stubs in place of all GUI
functions, and provides a library `solvespace_headless` alongside
the main executable. To cut down build times, only the few files
that have #if defined(HEADLESS) are built twice for the executable
and the library; the rest is grouped into a new `solvespace_cad`
library. It is not usable on its own and just serves for grouping.
This commit also gates the tests behind a -DENABLE_TESTS=ON CMake
option, ON by default (but suggested as OFF in the README so that
people don't ever have to install cairo to build the executable.)
The tests introduced in this commit are (so far) rudimentary,
although functional, and they serve as a stepping point towards
introducing coverage analysis.
This has several desirable consequences:
* It is now possible to port SolveSpace to a later version of
OpenGL, such as OpenGLES 2, so that it runs on platforms that
only have that OpenGL version;
* The majority of geometry is now rendered without references to
the camera in C++ code, so a renderer can now submit it to
the video card once and re-rasterize with a different projection
matrix every time the projection is changed, avoiding expensive
reuploads;
* The DOGD (draw or get distance) interface is now
a straightforward Canvas implementation;
* There are no more direct references to SS.GW.(projection)
in sketch rendering code, which allows rendering to multiple
viewports;
* There are no more unnecessary framebuffer flips on CPU on Cocoa
and GTK;
* The platform-dependent GL code is now confined to rendergl1.cpp.
* The Microsoft and Apple headers required by it that are prone to
identifier conflicts are no longer included globally;
* The rendergl1.cpp implementation can now be omitted from
compilation to run SolveSpace headless or with a different
OpenGL version.
Note these implementation details of Canvas:
* GetCamera currently always returns a reference to the field
`Camera camera;`. This is so that a future renderer that caches
geometry in the video memory can define it as asserting, which
would provide assurance against code that could accidentally
put something projection-dependent in the cache;
* Line and triangle rendering is specified through a level of
indirection, hStroke and hFill. This is so that a future renderer
that batches geometry could cheaply group identical styles.
* DrawPixmap and DrawVectorText accept a (o,u,v) and not a matrix.
This is so that a future renderer into an output format that
uses 2d transforms (e.g. SVG) could easily derive those.
Some additional internal changes were required to enable this:
* Pixmap is now always passed as std::shared_ptr<{const ,}Pixmap>.
This is so that the renderer could cache uploaded textures
between API calls, which requires it to capture a (weak)
reference.
* The PlatformPathEqual function was properly extracted into
platform-specific code. This is so that the <windows.h> header
could be included only where needed (in platform/w32* as well
as rendergl1.cpp).
* The SBsp{2,3}::DebugDraw functions were removed. They can be
rewritten using the Canvas API if they are ever needed.
While no visual changes were originally intended, some minor fixes
happened anyway:
* The "emphasis" yellow line from top-left corner is now correctly
rendered much wider.
* The marquee rectangle is now pixel grid aligned.
* The hidden entities now do not clobber the depth buffer, removing
some minor artifacts.
* The workplane "tab" now scales with the font used to render
the workplane name.
* The workplane name font is now taken from the normals style.
* Workplane and constraint line stipple is insignificantly
different. This is so that it can reuse the existing stipple
codepaths; rendering of workplanes and constraints predates
those.
Some debug functionality was added:
* In graphics window, an fps counter that becomes red when
rendering under 60fps is drawn.
This is to ensure that:
* it is clear, when looking at the point of usage, what is
the purpose of "true" or "false";
* when refactoring, a simple search will bring up any places that
need to be changed.
Also, argument names were synchronized between declaration and
implementation.
As an exception, these are not annotated:
* Printf(/*halfLine=*/), to avoid pointless churn.
Specifically, this enables -Wswitch=error on GCC/Clang and its MSVC
equivalent; the exact way it is handled varies slightly, but what
they all have in common is that in a switch statement over an
enumeration, any enumerand that is not explicitly (via case:) or
implicitly (via default:) handled in the switch triggers an error.
Moreover, we also change the switch statements in three ways:
* Switch statements that ought to be extended every time a new
enumerand is added (e.g. Entity::DrawOrGetDistance(), are changed
to explicitly list every single enumerand, and not have a
default: branch.
Note that the assertions are kept because it is legal for
a enumeration to have a value unlike any of its defined
enumerands, and we can e.g. read garbage from a file, or
an uninitialized variable. This requires some rearranging if
a default: branch is undesired.
* Switch statements that ought to only ever see a few select
enumerands, are changed to always assert in the default: branch.
* Switch statements that do something meaningful for a few
enumerands, and ignore everything else, are changed to do nothing
in a default: branch, under the assumption that changing them
every time an enumerand is added or removed would just result
in noise and catch no bugs.
This commit also removes the {Request,Entity,Constraint}::UNKNOWN and
Entity::DATUM_POINT enumerands, as those were just fancy names for
zeroes. They mess up switch exhaustiveness checks and most of the time
were not the best way to implement what they did anyway.
This follows the previous commit. Unlike it, though, a small change
to control flow is made to separate the command and pending operation
enumerations.
Specifically, take the old code that looks like this:
class Foo {
enum { X = 1, Y = 2 };
int kind;
}
... foo.kind = Foo::X; ...
and convert it to this:
class Foo {
enum class Kind : uint32_t { X = 1, Y = 2 };
Kind kind;
}
... foo.kind = Foo::Kind::X;
(In some cases the enumeration would not be in the class namespace,
such as when it is generally useful.)
The benefits are as follows:
* The type of the field gives a clear indication of intent, both
to humans and tools (such as binding generators).
* The compiler is able to automatically warn when a switch is not
exhaustive; but this is currently suppressed by the
default: ssassert(false, ...)
idiom.
* Integers and plain enums are weakly type checked: they implicitly
convert into each other. This can hide bugs where type conversion
is performed but not intended. Enum classes are strongly type
checked.
* Plain enums pollute parent namespaces; enum classes do not.
Almost every defined enum we have already has a kind of ad-hoc
namespacing via `NAMESPACE_`, which is now explicit.
* Plain enums do not have a well-defined ABI size, which is
important for bindings. Enum classes can have it, if specified.
We specify the base type for all enums as uint32_t, which is
a safe choice and allows us to not change the numeric values
of any variants.
This commit introduces absolutely no functional change to the code,
just renaming and change of types. It handles almost all cases,
except GraphicsWindow::pending.operation, which needs minor
functional change.
Grid fitting is performed only on glyph boundaries, since glyphs
include curves converted to pwl, which would be mangled by per-point
grid fitting.
Grid fitting is only performed when the plane in which text is
laid out is parallel to the viewing plane.
Grid fitting is only performed when rendering for display; there
are no devices with dpi low enough for grid fitting to become
profitable, and in any case we cannot predict what the dpi would
be anyway.
Before this change, the two buttons "Show/hide shaded model" (S) and
"Show/hide hidden lines" (H) resulted in drawing the following
elements in the following styles:
Button | Non-occluded | Non-occluded | Occluded | Occluded
state | solid edges | entities | solid edges | entities
--------+--------------+--------------+-------------+--------------
!S !H | | | solid-edge | entity style
--------+ | +-------------+--------------
S !H | | | invisible
--------+ solid-edge | entity style +-------------+--------------
!S H | | | |
--------+ | | solid-edge | entity style
S H | | | |
--------+--------------+--------------+-------------+--------------
After this change, they are drawn as follows:
Button | Non-occluded | Non-occluded | Occluded | Occluded
state | solid edges | entities | solid edges | entities
--------+--------------+--------------+-------------+--------------
!S !H | | | solid-edge | entity style
--------+ | +-------------+--------------
S !H | | | invisible
--------+ solid-edge | entity style +-------------+--------------
!S H | | | |
--------+ | | hidden-edge | stippled¹
S H | | | |
--------+--------------+--------------+-------------+--------------
¹ entity style, but the stipple parameters taken from hidden-edge
In SolveSpace's true WYSIWYG tradition, the 2d view export follows
the rendered view exactly.
Also, it is now possible to edit the stipple parameters of built-in
styles, so that by changing the hidden-edge style to non-stippled
it is possible to regain the old behavior.
This change is quite subtle. The goal is to improve responsiveness
of highlighting even further. To understand this change you need
to keep in mind that Windows and Gtk have dramatically different
behavior for paint (WM_PAINT in Windows, expose in Gtk) and
mouse move events.
In Windows, WM_PAINT and WM_MOUSEMOVE, unless sent explicitly,
are synthesized: WM_MOUSEMOVE is delivered when there are no other
messages and the current cursor position doesn't match the remembered
one, and WM_PAINT is delivered when there are no other messages,
even WM_MOUSEMOVE. This is pretty clever because it doesn't swamp
programs that are slow to process either of those events with even
more of them, ensuring they remain responsive.
In Gtk, expose events are delivered at the end of the frame whenever
there is an invalid view, and every single mouse move that happened
will result in a separate event.
If mouse move events are handled quickly, then the behavior is
identical in either case:
* process mouse move event
* perform hit testing
* invalidate view
* no more events to process!
* there are invalid views
* repaint
If, however, mouse move events are handled slower, then the behavior
diverges. With Gtk:
* process mouse move event
* perform hit testing (slow)
* while this happens, ten more mouse move events are added
* invalidate view
* end of frame!
* there are invalid views
* repaint
* process mouse move event...
As a result, the Gtk-hosted UI hopelessly lags behind user input.
This is very irritating.
With Windows:
* process mouse move event
* perform hit testing (slow)
* while this happens, mouse was moved
* invalidate view
* process mouse move event...
As a result, the Windows-hosted UI never repaints while the mouse
is moved. This is also very irritating.
Commit HEAD^ has fixed the problems with Gtk-based UI by making
hit testing so fast that mouse move events never quite overflow
the queue. There's still a barely noticeable lag but it's better.
However, the problems with Windows remained because while the queue
doesn't *overflow* with the faster hit testing code, it doesn't go
*empty* either! Thus we still don't repaint.
This commit builds on top of HEAD^ and makes it so that we don't
actually hit test anything if we haven't painted the result of
the previous hit test already. This fixes the problem on Windows
but also helps Gtk a little bit.
Curiously, the Cocoa-based UI never suffered from any of these
problems. To my understanding (it's somewhat underdocumented), it
processes mouse moves like Windows, but paints like Gtk.
This results in massive performance improvements for hit testing.
Files with very large amounts of entities (e.g. [1]) inflict
a delay of several seconds between moving the pointer and
highlighting an entity in commit HEAD^^^, whereas in this commit
the delay is barely perceptible.
[1]: http://solvespace.com/forum.pl?action=viewthread&parent=872
In my (whitequark's) experience this warning tends to expose
copy-paste errors with a high SNR, so making a few fragments
slightly less symmetric is worth it.
Also mollify -Wlogical-op-parentheses while we're at it.
After commit 2f734d9, inactive groups are no longer regenerated
for trivial changes, e.g. changing parameters, so it's possible to
switch to an earlier group and work on it without incurring
the computational (slowdown) and cognitive (annoyance by red
background) overhead of later groups failing to solve.
However, if a group--any group anywhere--was not solved OK,
the interface reacted accordingly, which diminished usefulness of
the change, especially given that, if we have groups A and B with
B depending on A, if B is broken by a change in A and we activate A
and fix it, B will not be regenerated.
After this commit, only active groups are considered when deciding
if generating the entire sketch would fail.
Before this commit, a single chord tolerance was used for both
displaying and exporting geometry. Moreover, this chord tolerance
was specified in screen pixels, and as such depended on zoom level.
This was inconvenient: exporting geometry with a required level of
precision required awkward manipulations of viewport. Moreover,
since some operations, e.g. mesh watertightness checking, were done
on triangle meshes which are generated differently depending on
the zoom level, these operations could report wildly different
and quite confusing results depending on zoom level.
The chord tolerance for display and export pursue completely distinct
goals: display chord tolerance should be set high enough to achieve
both fast regeneration and legible rendering, whereas export chord
tolerance should be set to match the dimension tolerance of
the fabrication process.
This commit introduces two distinct chord tolerances: a display
and an export one. Both chord tolerances are absolute and expressed
in millimeters; this is inappropriate for display purposes but
will be fixed in the next commits.
After exporting, the geometry is redrawn with the chord tolerance
configured for the export and an overlay message is displayed;
pressing Esc clears the message and returns the display back to
normal.
This will allow us to use non-POD classes inside these objects
in future and is otherwise functionally equivalent, as well
as more concise.
Note that there are some subtleties with handling of
brace-initialization. Specifically:
On aggregates (e.g. simple C-style structures) using an empty
brace-initializer zero-initializes the aggregate, i.e. it makes
all members zero.
On non-aggregates an empty brace-initializer calls the default
constructor. And if the constructor doesn't explicitly initialize
the members (which the auto-generated constructor doesn't) then
the members will be constructed but otherwise uninitialized.
So, what is an aggregate class? To quote the C++ standard
(C++03 8.5.1 §1):
An aggregate is an array or a class (clause 9) with no
user-declared constructors (12.1), no private or protected
non-static data members (clause 11), no base classes (clause 10),
and no virtual functions (10.3).
In SolveSpace, we only have to handle the case of base classes;
Constraint and Entity have those. Thus, they had to gain a default
constructor that does nothing but initializes the members to zero.
