That gives the change of storing that data in flash storage in some
architectures. That doesn't happen yet though, some extra changes are required
for that to happen.
We don't actually need all of them, since the second half is for the opposite
triangles. In that case we just need to negate the resulting vector when
changing basis.
That function was only being used by the unit tests and the benchmark. In order
to remove memory usage, the triangles will be removed, since we don't actually
need to keep them in real situations. Thus, this patch removes that function
and copy those triangles to the test and benchmark.
This is the second optimization. With that we don't have to iterate over the
umbrella's components.
The table below summarizes the mean CPU time in ns from the brenchmark results
on an Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz processor:
| Naive implementation | First Optimization | Second Optimization
------------------------------------------------------------------------
Min. | 26.0 | 28.00 | 26.0
1st Qu.| 78.0 | 48.75 | 39.0
Median | 132.0 | 57.00 | 41.0
Mean | 130.1 | 61.20 | 41.6
3rd Qu.| 182.2 | 76.00 | 47.0
Max. | 234.0 | 98.00 | 54.0
If v is the null vector, then alpha * v is still the null vector for any alpha
as a real number. That means that the null vector doesn't cross any section.
This is a first optimization of the algorithm. The struct for the neighbor
umbrella has only one member, but new members will be added in the next
optimization.
The table below summarizes the mean CPU time in ns from the brenchmark results
on an Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz processor:
Cases | Naive implementation | First Optimization
--------------------------------------------------
Min. | 26.0 | 28.00
1st Qu.| 78.0 | 48.75
Median | 132.0 | 57.00
Mean | 130.1 | 61.20
3rd Qu.| 182.2 | 76.00
Max. | 234.0 | 98.00
This optimization reduces the mean time for the worst case (Max. line) by more
than 50%.
- Change the order of the icosahedron triangles so that there's a uniform way of
finding the opposite triangle. The order visually still makes sense.
- Change test to accommodate the order change.
Avoid warnings like:
[2130/2168] Compiling libraries/AP_Math/tests/test_math.cpp
../../libraries/AP_Math/tests/test_math.cpp: In member function ‘virtual void MathTest_IsZero_Test::TestBody()’:
../../libraries/AP_Math/tests/test_math.cpp:73:196: warning: converting ‘false’ to pointer type for argument 1 of ‘char
testing::internal::IsNullLiteralHelper(testing::internal::Secret*)’ [-Wconversion-null]
../../libraries/AP_Math/tests/test_math.cpp:74:199: warning: converting ‘false’ to pointer type for argument 1 of ‘char
testing::internal::IsNullLiteralHelper(testing::internal::Secret*)’ [-Wconversion-null]
Use EXPECT_TRUE() and EXPECT_FALSE() from gtest instead.
The new function can deal with a variable number of function parameters.
Additionally, I renamed the functions to norm(), because this is the
standard name used in several other projects.
When using wrap_180_cd() we are adding a small float (180 * 100) to a
possibly big number. This may lose float precision as illustrated by the
unit test failing:
OUT: ../../libraries/AP_Math/tests/test_math.cpp:195: Failure
OUT: Value of: wrap_180_cd(-3600000000.f)
OUT: Actual: -80
OUT: Expected: 0.f
OUT: Which is: 0
These functions (or variants thereof) now have unit tests:
- is_zero()
- is_equal()
- sq()
- pythagorous()
- constrain()
- wrap_180()
- wrap_360()
Some tests in wrap_180_cd are failing: -180 should be wrapped to 180,
not -180.
For ROTATION_ROLL_90_PITCH_68_YAW_293 consider the angles as 90, 68.8
and 293.3 degrees to pre-calculate rotation. This matches the rotation
matrix used in code.
While at it, check not only the values are close enough but also the
length of the vector.
The rotations are supposed to follow the name of the enum, in order. The
ROTATION_YAW_293_PITCH_68_ROLL_90 was added with the name of an
intrinsic 321 rotation, but the matrix is actually a 123 rotation,
following the other rotations already present.
Change the name to follow the other names.
The problem with the current MIN/MAX macros is that they evaluate twice
the arguments. For example, these cases provide unintended results:
// a is incremented twice
a = MIN(a++, b);
// foo() with side-effects
a = MIN(foo(), b);
The alternative implementation here was provided by Daniel Frenzel using
template function. It doesn't have type safety as std::min and std::max,
but adding type safety would mean to check case by case what would be a
reasonable type and add proper casts. Here the arguments for MIN and MAX
can have different types and the return type is deduced from the
expression in the function.
Inspecting the current callers no place was found with the unintended
results above, but some in which now we don't calculate twice the
parameters will benefit from this new version. Examples:
float velocity_max = MIN(_pos_control.get_speed_xy(), safe_sqrt(0.5f*_pos_control.get_accel_xy()*_radius));
float acro_level_mix = constrain_float(1-MAX(MAX(abs(roll_in), abs(pitch_in)), abs(yaw_in))/4500.0, 0, 1)*ahrs.cos_pitch()
accel_x_cmss = (GRAVITY_MSS * 100) * (-(_ahrs.cos_yaw() * _ahrs.sin_pitch() / MAX(_ahrs.cos_pitch(),0.5f)) - _ahrs.sin_yaw() * _ahrs.sin_roll() / MAX(_ahrs.cos_roll(),0.5f));
track_leash_slack = MIN(_track_leash_length*(leash_z-track_error_z)/leash_z, _track_leash_length*(leash_xy-track_error_xy)/leash_xy);
RC_Channel_aux::move_servo(RC_Channel_aux::k_sprayer_pump, MIN(MAX(ground_speed * _pump_pct_1ms, 100 *_pump_min_pct),10000),0,10000);
The problem with using min() and max() is that they conflict with some
C++ headers. Name the macros in uppercase instead. We may go case by
case later converting them to be typesafe.
Changes generated with:
git ls-files '*.cpp' '*.h' -z | xargs -0 sed -i 's/\([^_[:alnum:]]\)max(/\1MAX(/g'
git ls-files '*.cpp' '*.h' -z | xargs -0 sed -i 's/\([^_[:alnum:]]\)min(/\1MIN(/g'
Remove the checks for HAL_CPU_CLASS > HAL_CPU_CLASS_16 and
HAL_CPU_CLASS >= HAL_CPU_CLASS_75. Corresponding dead code will be
removed on separate commits.
Most of AP_Progmem is already gone so we can stop including it in most
of the places. The only places that need it are the ones using
pgm_read_*() APIs.
In some cases the header needed to be added in the .cpp since it was
removed from the .h to reduce scope. In those cases the headers were
also reordered.
Now variables don't have to be declared with PROGMEM anymore, so remove
them. This was automated with:
git grep -l -z PROGMEM | xargs -0 sed -i 's/ PROGMEM / /g'
git grep -l -z PROGMEM | xargs -0 sed -i 's/PROGMEM//g'
The 2 commands were done so we don't leave behind spurious spaces.
AVR-specific places were not changed.
The PSTR is already define as a NOP for all supported platforms. It's
only needed for AVR so here we remove all the uses throughout the
codebase.
This was automated with a simple python script so it also converts
places which spans to multiple lines, removing the matching parentheses.
AVR-specific places were not changed.
Remove unnecessary includes, in particular the includes for specific
boards. The list of libraries for 'polygon' example was updated so that
the example compiles again.
Instead of requiring every program to specify the HAL related modules,
let the build system do it (in practice everything we compiled depended
on HAL anyway). This allow including only the necessary files in the
compilation.
The switching between different AP_HAL was happening by giving different
definitions of AP_HAL_BOARD_DRIVER, and the programs would use it to
instantiate.
A program or library code would have to explicitly include (and depend)
on the concrete implementation of the HAL, even when using it only via
interface.
The proposed change move this dependency to be link time. There is a
AP_HAL::get_HAL() function that is used by the client code. Each
implementation of HAL provides its own definition of this function,
returning the appropriate concrete instance.
Since this replaces the job of AP_HAL_BOARD_DRIVER, the definition was
removed.
The static variables for PX4 and VRBRAIN were named differently to avoid
shadowing the extern symbol 'hal'.
compile time optimization
Functions like sq() are better moved to the header file as inline.
Compiler can then optimize these out when used in code, this saves cpu
cycles with stack push, pop during function calls.
This commit changes the way libraries headers are included in source files:
- If the header is in the same directory the source belongs to, so the
notation '#include ""' is used with the path relative to the directory
containing the source.
- If the header is outside the directory containing the source, then we use
the notation '#include <>' with the path relative to libraries folder.
Some of the advantages of such approach:
- Only one search path for libraries headers.
- OSs like Windows may have a better lookup time.
This commit changes the way libraries headers are included in source files:
- If the header is in the same directory the source belongs to, so the
notation '#include ""' is used with the path relative to the directory
containing the source.
- If the header is outside the directory containing the source, then we use
the notation '#include <>' with the path relative to libraries folder.
Some of the advantages of such approach:
- Only one search path for libraries headers.
- OSs like Windows may have a better lookup time.
per Randy's suggestion, fast_atan2() is no longer necessary over atan2() because only copter uses it and copter is no longer supported on future builds of APM
ccd578664f (commitcomment-11025083)
Gustavo Jose de Sousa
dgrat
Ricardo de Almeida Gonzaga
Lucas De Marchi
Andrew Tridgell
Michael du Breuil
Tom Pittenger
Daniel Frenzel
Peter Barker
Jonathan Challinger
bugobliterator
José Roberto de Souza
Caio Marcelo de Oliveira Filho
Randy Mackay
Valmantas Palikša