## Summary of the Pull Request
In the original implementation, we used two different orderings for the color tables. The WT color table used ANSI order, while the conhost color table used a Windows-specific order. This PR standardizes on the ANSI color order everywhere, so the usage of indexed colors is consistent across both parts of the code base, which will hopefully allow more of the code to be shared one day.
## References
This is another small step towards de-duplicating `AdaptDispatch` and `TerminalDispatch` for issue #3849, and is essentially a followup to the SGR dispatch refactoring in PR #6728.
## PR Checklist
* [x] Closes#11461
* [x] CLA signed.
* [x] Tests added/passed
* [ ] Documentation updated.
* [ ] Schema updated.
* [x] I've discussed this with core contributors already. Issue number where discussion took place: #11461
## Detailed Description of the Pull Request / Additional comments
Conhost still needs to deal with legacy attributes using Windows color order, so those values now need to be transposed to ANSI colors order when creating a `TextAttribute` object. This is done with a simple mapping table, which also handles the translation of the default color entries, so it's actually slightly faster than the original code.
And when converting `TextAttribute` values back to legacy console attributes, we were already using a mapping table to handle the narrowing of 256-color values down to 16 colors, so we just needed to adjust that table to account for the translation from ANSI to Windows, and then could make use of the same table for both 256-color and 16-color values.
There are also a few places in conhost that read from or write to the color tables, and those now need to transpose the index values. I've addressed this by creating separate `SetLegacyColorTableEntry` and `GetLegacyColorTableEntry` methods in the `Settings` class which take care of the mapping, so it's now clearer in which cases the code is dealing with legacy values, and which are ANSI values.
These methods are used in the `SetConsoleScreenBufferInfoEx` and `GetConsoleScreenBufferInfoEx` APIs, as well as a few place where color preferences are handled (the registry, shortcut links, and the properties dialog), none of which are particularly sensitive to performance. However, we also use the legacy table when looking up the default colors for rendering (which happens a lot), so I've refactored that code so the default color calculations now only occur once per frame.
The plus side of all of this is that the VT code doesn't need to do the index translation anymore, so we can finally get rid of all the calls to `XTermToWindowsIndex`, and we no longer need a separate color table initialization method for conhost, so I was able to merge a number of color initialization methods into one. We also no longer need to translate from legacy values to ANSI when generating VT sequences for conpty.
The one exception to that is the 16-color VT renderer, which uses the `TextColor::GetLegacyIndex` method to approximate 16-color equivalents for RGB and 256-color values. Since that method returns a legacy index, it still needs to be translated to ANSI before it can be used in a VT sequence. But this should be no worse than it was before.
One more special case is conhost's secret _Color Selection_ feature. That uses `Ctrl`+Number and `Alt`+Number key sequences to highlight parts of the buffer, and the mapping from number to color is based on the Windows color order. So that mapping now needs to be transposed, but that's also not performance sensitive.
The only thing that I haven't bothered to update is the trace logging code in the `Telemetry` class, which logs the first 16 entries in the color table. Those entries are now going to be in a different order, but I didn't think that would be of great concern to anyone.
## Validation Steps Performed
A lot of unit tests needed to be updated to use ANSI color constants when setting indexed colors, where before they might have been expecting values in Windows order. But this replaced a wild mix of different constants, sometimes having to use bit shifting, as well as values mapped with `XTermToWindowsIndex`, so I think the tests are a whole lot clearer now. Only a few cases have been left with literal numbers where that seemed more appropriate.
In addition to getting the unit tests working, I've also manually tested the behaviour of all the console APIs which I thought could be affected by these changes, and confirmed that they produced the same results in the new code as they did in the original implementation.
This includes:
- `WriteConsoleOutput`
- `ReadConsoleOutput`
- `SetConsoleTextAttribute` with `WriteConsoleOutputCharacter`
- `FillConsoleOutputAttribute` and `FillConsoleOutputCharacter`
- `ScrollConsoleScreenBuffer`
- `GetConsoleScreenBufferInfo`
- `GetConsoleScreenBufferInfoEx`
- `SetConsoleScreenBufferInfoEx`
I've also manually tested changing colors via the console properties menu, the registry, and shortcut links, including setting default colors and popup colors. And I've tested that the "Quirks Mode" is still working as expected in PowerShell.
In terms of performance, I wrote a little test app that filled a 80x9999 buffer with random color combinations using `WriteConsoleOutput`, which I figured was likely to be the most performance sensitive call, and I think it now actually performs slightly better than the original implementation.
I've also tested similar code - just filling the visible window - with SGR VT sequences of various types, and the performance seems about the same as it was before.
This commit enables /fp:fast. This doubles the performance of the Delta E
computation in #11095 for instance. Additionally it re-enables two options for
debug builds which are normally enabled by default by Visual Studio.
## PR Checklist
* [x] I work here
* [x] Tests added/passed
## Validation Steps Performed
* No change in binary size
* No obvious change in behavior
Shortly before adding the SSE2 variant I "improved" it by using
`_mm_packs_epi32`, but failed to test it again afterwards.
## PR Checklist
* [x] Closes#10866
* [x] I work here
## Validation Steps Performed
* `printf "\e[mNORMAL \e[1mBOLD\n"` results in correct bold white glyphs ✔️
I was watching a video about vectorized instructions and I wanted to
try out some new things, as I had never written AVX code before.
This commit is the result of this tiny Thursday morning detour into
AVX land. It improves performance of `TextColor::GetColor` by about 3x.
## Validation Steps Performed
* Default colors are still properly shifted +8 ✔️
The recent changes to use gsl::span everywhere added a few bounds checks
along codepaths where we were already checking bounds. Some of them may
be non-obvious to the optimizer, so we can now use til::at to help them
along.
To accomplish this, I've added a new overload of til::at that takes a
span and directly accesses its backing buffer.
We were using std::basic_string_view as a stand-in for std::span so that
we could change over all at once when C++20 dropped with full span
support. That day's not here yet, but as of 54a7fce3e we're using GSL 3,
whose span is C++20-compliant.
This commit replaces every instance of basic_string_view that was not
referring to an actual string with a span of the appropriate type.
I moved the `const` qualifier into span's `T` because while
`basic_string_view.at()` returns `const T&`, `span.at()` returns `T&`
(without the const). I wanted to maintain the invariant that members of
the span were immutable.
* Mechanical Changes
* `sv.at(x)` -> `gsl::at(sp, x)`
* `sv.c{begin,end}` -> `sp.{begin,end}` (span's iterators are const)
I had to replace a `std::basic_string<>` with a `std::vector<>` in
ConImeInfo, and I chose to replace a manual array walk in
ScreenInfoUiaProviderBase with a ranged-for. Please review those
specifically.
This will almost certainly cause a code size regression in Windows
because I'm blowing out all the PGO counts. Whoops.
Related: #3956, #975.
This PR reimplements the VT rendering engines to do a better job of
preserving the original color types when propagating attributes over
ConPTY. For the 16-color renderers it provides better support for
default colors and improves the efficiency of the color narrowing
conversions. It also fixes problems with the ordering of character
renditions that could result in attributes being dropped.
Originally the base renderer would calculate the RGB color values and
legacy/extended attributes up front, passing that data on to the active
engine's `UpdateDrawingBrushes` method. With this new implementation,
the renderer now just passes through the original `TextAttribute` along
with an `IRenderData` interface, and leaves it to the engines to extract
the information they need.
The GDI and DirectX engines now have to lookup the RGB colors themselves
(via simple `IRenderData` calls), but have no need for the other
attributes. The VT engines extract the information that they need from
the `TextAttribute`, instead of having to reverse engineer it from
`COLORREF`s.
The process for the 256-color Xterm engine starts with a check for
default colors. If both foreground and background are default, it
outputs a SGR 0 reset, and clears the `_lastTextAttribute` completely to
make sure any reset state is reapplied. With that out the way, the
foreground and background are updated (if changed) in one of 4 ways.
They can either be a default value (SGR 39 and 49), a 16-color index
(using ANSI or AIX sequences), a 256-color index, or a 24-bit RGB value
(both using SGR 38 and 48 sequences).
Then once the colors are accounted for, there is a separate step that
handles the character rendition attributes (bold, italics, underline,
etc.) This step must come _after_ the color sequences, in case a SGR
reset is required, which would otherwise have cleared any character
rendition attributes if it came last (which is what happened in the
original implementation).
The process for the 16-color engines is a little different. The target
client in this case (Windows telnet) is incapable of setting default
colors individually, so we need to output an SGR 0 reset if _either_
color has changed to default. With that out the way, we use the
`TextColor::GetLegacyIndex` method to obtain an approximate 16-color
index for each color, and apply the bold attribute by brightening the
foreground index (setting bit 8) if the color type permits that.
However, since Windows telnet only supports the 8 basic ANSI colors, the
best we can do for bright colors is to output an SGR 1 attribute to get
a bright foreground. There is nothing we can do about a bright
background, so after that we just have to drop the high bit from the
colors. If the resulting index values have changed from what they were
before, we then output ANSI 8-color SGR sequences to update them.
As with the 256-color engine, there is also a final step to handle the
character rendition attributes. But in this case, the only supported
attributes are underline and reversed video.
Since the VT engines no longer depend on the active color table and
default color values, there was quite a lot of code that could now be
removed. This included the `IDefaultColorProvider` interface and
implementations, the `Find(Nearest)TableIndex` functions, and also the
associated HLS conversion and difference calculations.
VALIDATION
Other than simple API parameter changes, the majority of updates
required in the unit tests were to correct assumptions about the way the
colors should be rendered, which were the source of the narrowing bugs
this PR was trying to fix. Like passing white on black to the
`UpdateDrawingBrushes` API, and expecting it to output the default `SGR
0` sequence, or passing an RGB color and expecting an indexed SGR
sequence.
In addition to that, I've added some VT renderer tests to make sure the
rendition attributes (bold, underline, etc) are correctly retained when
a default color update causes an `SGR 0` sequence to be generated (the
source of bug #3076). And I've extended the VT renderer color tests
(both 256-color and 16-color) to make sure we're covering all of the
different color types (default, RGB, and both forms of indexed colors).
I've also tried to manually verify that all of the test cases in the
linked bug reports (and their associated duplicates) are now fixed when
this PR is applied.
Closes#2661Closes#3076Closes#3717Closes#5384Closes#5864
This is only a partial fix for #293, but I suspect the remaining cases
are unfixable.
This PR provides a faster algorithm for converting 8-bit and 24-bit
colors into the 4-bit legacy values that are required by the Win32
console APIs. It also fixes areas of the code that were incorrectly
using a simple 16-color conversion that didn't handle 8-bit and 24-bit
values.
The faster conversion algorithm should be an improvement for issues #783
and #3950.
One of the main points of this PR was to fix the
`ReadConsoleOutputAttribute` API, which was using a simplified legacy
color conversion (the original `TextAttribute:GetLegacyAttributes`
method), which could only handle values from the 16-color table. RGB
values, and colors from the 256-color table, would be mapped to
completely nonsensical values. This API has now been updated to use the
more correct `Settings::GenerateLegacyAttributes` method.
But there were also a couple of other places in the code that were using
`GetLegacyAttributes` when they really had no reason to be working with
legacy attributes at all. This could result in colors being downgraded
to 4-bit values (often badly, as explained above), when the code was
already perfectly capable of displaying the full 24-bits.
This included the fill colors in the IME composer (in `ConsoleImeInfo`),
and the construction of the highlighting colors in the color
search/selection handler (`Selection::_HandleColorSelection`). I also
got rid of some legacy attribute code in the `Popup` class, which was
originally intended to update colors below the popup when the settings
changed, but actually caused more problems than it solved.
The other major goal of this PR was to improve the performance of the
`GenerateLegacyAttributes` method, since the existing implementation
could be quite slow when dealing with RGB values.
The simple cases are handled much the same as they were before. For an
`IsDefault` color, we get the default index from the
`Settings::_wFillAttribute` field. For an `IsIndex16` color, the index
can just be returned as is.
For an `IsRgb` color, the RGB components are compressed down to 8 bits
(3 red, 3 green, 2 blue), simply by dropping the least significant bits.
This 8-bit value is then used to lookup a representative 16-color value
from a hard-coded table. An `IsIndex256` color is also converted with a
lookup table, just using the existing 8-bit index.
The RGB mapping table was calculated by taking each compressed 8-bit
color, and picking a entry from the _Campbell_ palette that best
approximated that color. This was done by looking at a range of 24-bit
colors that mapped to the 8-bit value, finding the best _Campbell_ match
for each of them (using a [CIEDE2000] color difference calculation), and
then the most common match became the index that the 8-bit value would
map to.
The 256-color table was just a simpler version of this process. For each
entry in the table, we take the default RGB palette value, and find it's
closest match in the _Campbell_ palette.
Because these tables are hard-coded, the results won't adjust to changes
in the palette. However, they should still produce reasonable results
for palettes that follow the standard ANSI color range. And since
they're only a very loose approximation of the colors anyway, the exact
value really isn't that important.
That said, I have tried to make sure that if you take an RGB value for a
particular index in a reasonable color scheme, then the legacy color
mapped from that value should ideally match the same index. This will
never be possible for all color schemes, but I have tweaked a few of the
table entries to improve the results for some of the common schemes.
One other point worth making regarding the hard-coded tables: even if we
wanted to take the active palette into account, that wouldn't actually
be possible over a conpty connection, because we can't easily know what
color scheme the client application is using. At least this way the
results in conhost are guaranteed to be the same as in the Windows
Terminal.
[CIEDE2000]: https://en.wikipedia.org/wiki/Color_difference#CIEDE2000
## Validation Steps Performed
This code still passes the `TextAttributeTests` that check the basic
`GetLegacyAttribute` behaviour and verify the all legacy attributes
roundtrip correctly. However, some of the values in the `RgbColorTests`
had to be updated, since we're now intentionally returning different
values as a result of the changes to the RGB conversion algorithm.
I haven't added additional unit tests, but I have done a lot of manual
testing to see how well the new algorithm works with a range of colors
and a variety of different color schemes. It's not perfect in every
situation, but I think it works well enough for the purpose it serves.
I've also confirmed that the issues reported in #5940 and #6247 are now
fixed by these changes.
Closes#5940Closes#6247
This PR introduces a new `ColorType` to allow us to distinguish between
`SGR` indexed colors from the 16 color table, the lower half of which
can be brightened, and the ISO/ITU indexed colors from the 256 color
table, which have a fixed brightness. Retaining the distinction between
these two types will enable us to forward the correct `SGR` sequences to
conpty when addressing issue #2661.
The other benefit of retaining the color index (which we didn't
previously do for ISO/ITU colors) is that it ensures that the colors are
updated correctly when the color scheme is changed.
## References
* This is another step towards fixing the conpty narrowing bugs in issue
#2661.
* This is technically a fix for issue #5384, but that won't be apparent
until #2661 is complete.
## PR Checklist
* [x] Closes#1223
* [x] CLA signed.
* [x] Tests added/passed
* [ ] Requires documentation to be updated
* [x] I've discussed this with core contributors already.
## Detailed Description of the Pull Request / Additional comments
The first part of this PR was the introduction of a new `ColorType` in
the `TextColor` class. Instead of just the one `IsIndex` type, there is
now an `IsIndex16` and an `IsIndex256`. `IsIndex16` covers the eight
original ANSI colors set with `SGR 3x` and `SGR 4x`, as well as the
brighter aixterm variants set with `SGR 9x` and `SGR 10x`. `IsIndex256`
covers the 256 ISO/ITU indexed colors set with `SGR 38;5` and `SGR
48;5`.
There are two reasons for this distinction. The first is that the ANSI
colors have the potential to be brightened by the `SGR 1` bold
attribute, while the ISO/ITO color do not. The second reason is that
when forwarding an attributes through conpty, we want to try and
preserve the original SGR sequence that generated each color (to the
extent that that is possible). By having the two separate types, we can
map the `IsIndex16` colors back to ANSI/aixterm values, and `IsIndex256`
to the ISO/ITU sequences.
In addition to the VT colors, we also have to deal with the legacy
colors set by the Windows console APIs, but we don't really need a
separate type for those. It seemed most appropriate to me to store them
as `IsIndex256` colors, since it doesn't make sense to have them
brightened by the `SGR 1` attribute (which is what would happen if they
were stored as `IsIndex16`). If a console app wanted a bright color it
would have selected one, so we shouldn't be messing with that choice.
The second part of the PR was the unification of the two color tables.
Originally we had a 16 color table for the legacy colors, and a separate
table for the 256 ISO/ITU colors. These have now been merged into one,
so color table lookups no longer need to decide which of the two tables
they should be referencing. I've also updated all the methods that took
a color table as a parameter to use a `basic_string_view` instead of
separate pointer and length variables, which I think makes them a lot
easier and safer to work with.
With this new architecture in place, I could now update the
`AdaptDispatch` SGR implementation to store the ISO/ITU indexed colors
as `IsIndex256` values, where before they were mapped to RGB values
(which prevented them reflecting any color scheme changes). I could also
update the `TerminalDispatch` implementation to differentiate between
the two index types, so that the `SGR 1` brightening would only be
applied to the ANSI colors.
I've also done a bit of code refactoring to try and minimise any direct
access to the color tables, getting rid of a lot of places that were
copying tables with `memmove` operations. I'm hoping this will make it
easier for us to update the code in the future if we want to reorder the
table entries (which is likely a requirement for unifying the
`AdaptDispatch` and `TerminalDispatch` implementations).
## Validation Steps Performed
For testing, I've just updated the existing unit tests to account for
the API changes. The `TextColorTests` required an extra parameter
specifying the index type when setting an index. And the `AdapterTest`
and `ScreenBufferTests` required the use of the new `SetIndexedXXX`
methods in order to be explicit about the index type, instead of relying
on the `TextAttribute` constructor and the old `SetForeground` and
`SetBackground` methods which didn't have a way to differentiate index
types.
I've manually tested the various console APIs
(`SetConsoleTextAttribute`, `ReadConsoleOutputAttribute`, and
`ReadConsoleOutput`), to make sure they are still setting and reading
the attributes as well as they used to. And I've tested the
`SetConsoleScreenBufferInfoEx` and `GetConsoleScreenBufferInfoEx` APIs
to make sure they can read and write the color table correctly. I've
also tested the color table in the properties dialog, made sure it was
saved and restored from the registry correctly, and similarly saved and
restored from a shortcut link.
Note that there are still a bunch of issues with the color table APIs,
but no new problems have been introduced by the changes in this PR, as
far as I could tell.
I've also done a bunch of manual tests of `OSC 4` to make sure it's
updating all the colors correctly (at least in conhost), and confirmed
that the test case in issue #1223 now works as expected.
Generated by https://github.com/jsoref/spelling `f`; to maintain your repo, please consider `fchurn`
I generally try to ignore upstream bits. I've accidentally included some items from the `deps/` directory. I expect someone will give me a list of items to drop, I'm happy to drop whole files/directories, or to split the PR into multiple items (E.g. comments/locals/public).
Closes#4294
