Part 1 of the big GPU merge #45

bcumming · 2016-10-27T12:43:56Z

This PR is part of the gpu feature merge. The GPU implementation is not implemented here. Instead, we focus on refactoring of the original "multicore" back end so that it is ready for adding the GPU back end.

This is a big and messy change, for which I am sorry.

build System

A WITH_CUDA option has been added to the main CMakeLists. This finds the CUDA toolkit, and sets CUDA compiler flags, and will build unit tests for the gpu back end.
The CMakeLists that generates mechanisms with modcc has been updated to generate CUDA mechanisms.
the library is now named libnestmc instead of libcellalgo
merge the external libraries that are optionally linked againts (tbb, libunwind, etc) into a single
EXTERNAL_LIBRARIES list for ease of linking

modcc

the cprinter and cudaprinter have had small changes to generate mechanism files that are compatible with the refactored library.

algorithms

the indexes into algorithm was "rangified". An algorithm index_into_iterator takes two ranges as inputs to make a range that lazily generates the index of sub into super set.

backends

made a new path src/backends/ for backend specific type and implementation code.
currently:
- complete support for the multicore and gpu backends
- gpu back end is not optimized or validated
the back end implementations are in src/backends
a single backend class, nest::mc::{multicore,gpu}::backend, is provides all backend specific type and implementation details fro each backend
- storage containers
- Hines matrix assembly for FVM method
- Hines matrix solver
- mechanism "factory"

lowered fvm cells

removed fvm_cell because this can be modelled with an fvm_multicell with one cell.
refactored to use backend type and implementation from fvm_policy
use std::vector instead of containers in nest::mc::memory:: where possible when building cells.

memory library

Refactor the "memory" library, making it much simpler and better integrated into the rest of the application. However, it is still far from perfect. The Coordinator approach needs to be improved, most likely by putting target-specific wisdom into pointers (which could obviate the need for a const_array_view type.

renaming and moving
- move from vector/ to src/memory
- move into the nest::mc namespace, i.e. all types and functions are now in nest::mc::memory
- change from camel case nameing scheme to NestMC style naming.
simplification
- remove the CRTP cruft that was used to make operator overloading work for operations like copying from one range into another, and filling a range with a constant value. These have been replaced with memory::fill() and memory::copy() helper functions. This simplified the code a lot, and makes code clearer in user land.
```
// before
vec(0, 5) = other;
// now
memory::copy(other, vec(0, 5));
```
- add some wrappers in src/memory/wrappers.hpp that help with making views. These are particularly useful for passing std::vector through interfaces that expect a view.

debug backtraces

Added stack traces for debugging.

support for OSX and Linux via libunwind
backtraces can be generated manually nest::mc::util::backtrace().print()
- creates a new file and dumps trace into file
- prints message to stderr with file name and instructions on how to analyse
backtraces are also automatically generated when an assertion EXPECTS statement fails
a python script in scripts/print_backtrace pretty prints the output with file name, line number and demangled symbols

util simplification and consolidation

The src/util.hpp file was removed

much of its contents were dead code and just removed
useful components like pprintf and make_unique were moved into the src/utils path in standalone files
There was a lot of overlap between functionality provided in src/memory/util.hpp and existing functions/types in the nest::mc::util namespace. The memory implementations were removed, and their nest::mc::util counterparts used. There is still some work remaining, namely moving the rest of the src/memory/util.hpp into src/util/...

* add CMake option and target for CUDA * add some workarounds for CUDA compiler bugs

* created a gpu multicell target type * missing CUDA specific kernels * basic cell state (voltage, current, etc) is on gpu * moved the vector library into main source tree * have not yet deleted external dependency

* more simplifications of the memory library * clean up of policy based matrix gpu/multicore implementation

snapshot of code before attempting to merge the multicore and gpu implementation of the FVM into a single code base. The gpu version of the miniapp doesn't compile, but fixing this is better done as part of merging. * fixed bug in multicore fvm_multicell where solutions wasn't copied from rhs to voltage vector after linear system solution * add check for ion channel CV indexes to unit tests - this fails due to CV bug that is in issue 20 * small fixes towards gpu support

merge from main

* Add documentation of template parameters for `filter_iterator`. * Document use of `uninitalized<F>` for holding functional objects in `filter_iterator` and `transform_iterator`

More range functionality, unit tests.

Feature/new test model

* Simplify trace analysis and reporting code in `trace_analysis.hpp` * Consolidate convergence test run procedures into new class `convergence_test_runner`.

* Make `algorithm::sum`, `algorithm::mean` more generic, allowing use with array types. * Add `div_compartment` compartment representation, that holds geometric information for each half of a compartment that will then be used in calculating control volumes. * Add three compartmentalisation schemes/policies that discretize a segment into `div_compartment` objects: * `div_compartment_by_ends` divides based only on the segment end points and radii. * `div_compartment_sampler` forms frusta by sampling the segment radius at each compartment boundary * `div_compartment_integrator` computes the compartment areas and volumes exactly by summing all frustra in the intersection of the segment and the compartmnet span.

…ests Consolidate validation tests (issue #41)

Feature/divided compartments

halfflat · 2016-10-28T09:38:19Z

CMakeLists.txt

+if(WITH_CUDA)
+    find_package(CUDA REQUIRED)
+    # the vector library has a compiled component when using the CUDA backend
+    include(ExternalProject)


No longer need ExternalProject here I think.

halfflat · 2016-10-28T09:58:49Z

miniapp/CMakeLists.txt

-target_link_libraries(miniapp.exe LINK_PUBLIC cellalgo)
-target_link_libraries(miniapp.exe LINK_PUBLIC ${TBB_LIBRARIES})
+if(WITH_TBB)
+    target_link_libraries(miniapp.exe ${TBB_LIBRARIES})


Need a LINK_PUBLIC in here and on line 16 in order to avoid CMP0023 nonsense.

halfflat · 2016-10-28T10:05:12Z

modcc/cprinter.cpp

-    text_.add_line("data_ = vector_type(field_size * num_fields);");
-    text_.add_line("data_(memory::all) = std::numeric_limits<value_type>::quiet_NaN();");
+    text_.add_line("data_ = array(field_size * num_fields);");
+    text_.add_line("memory::fill(data_, std::numeric_limits<value_type>::quiet_NaN());");


Is it worth considering an STL-like constructor for array which can take a fill value? c.f. std::vector<T>(size_type count, const T& value = T())

The library already has this functionality. I don't know why I have done things this way, probably just sloppiness.

Yep, it was sloppiness. I have fixed it by using just one line:

text_.add_line("data_ = array(field_size*num_fields, std::numeric_limits<value_type>::quiet_NaN());");

halfflat · 2016-10-28T11:03:59Z

src/algorithms.hpp

+    using sub_iterator = typename R2::const_iterator;
+
+    mutable super_iterator super_it_;
+    const super_iterator super_end_;


In principle, the super range may be delimited by a sentinel of different type to super_iterator. Rather than being parameterized on the two sequence types R1 and R2, I think it would be cleaner to parameterize it on the sub iterator type, the super iterator type, and the super sentinel type.

I agree.

index_into_iterator is now templated on the 3 iterator types

updated unit test to check that arrays work specifically

halfflat · 2016-10-28T11:06:28Z

src/algorithms.hpp

+    public std::iterator<std::forward_iterator_tag, typename R1::value_type>
+{
+public:
+    using value_type = typename R1::value_type;


As dereference returns an index (offset into super range), it should be e.g. typename std::iterator_traits<super_iterator>::difference_type or typename util::sequence_traits<R1>::size_type.

do you mean

typename util::sequence_traits<R1>::difference_type

instead of size_type?

If the class is ultimately going to be defined over iterator types rather than range types, then iterator_traits<super_iterator>::difference_type will be the way to go. Analogous to indices for e.g. std::string, an unsigned type might be preferable.

halfflat · 2016-10-28T12:57:37Z

src/memory/array_view.hpp

+} // namespace memory
+} // namespace mc
+} // namespace nest
+


(I'm basically trusting these are just lightly renamed and moved versions of the original Vector code.)

yes, they are.

halfflat · 2016-10-28T13:01:00Z

src/memory/gpu.hpp

+        bytes, cudaMemcpyDeviceToHost
+    );
+    if(status != cudaSuccess) {
+        LOG_ERROR("cudaMemcpy(d2h, " + std::to_string(bytes) + ") " + cudaGetErrorString(status));


We've now got multiple error, verbosity macros in use under src/; these should be consolidated in one place with consistent usage across the code.

I agree! Let's put it in the code cleanup bucket.

Commenting just to test slack trigger.

halfflat · 2016-10-28T13:05:17Z

src/spike_source.hpp

@@ -12,7 +12,8 @@ class spike_detector {
 public:
    using cell_type = Cell;

-    spike_detector(const cell_type& cell, typename Cell::detector_handle h, double thresh, float t_init) :
+    spike_detector(
+        const cell_type& cell, typename Cell::detector_handle h, double thresh, float t_init) :


Parameters line up with initializers, without any intervening punctuation. I think it's a confusing compromise. We can nut this out on our coding guidelines bike shed channel.

changed to the following... not perfect either, but at least it is clear to the reader what is going on (IMO)

spike_detector( ~ const cell_type& cell, + typename Cell::detector_handle h, + double thresh, + float t_init + ): handle_(h), threshold_(thresh) { reset(cell, t_init); }

halfflat · 2016-10-28T13:14:19Z

src/util/rangeutil.hpp

+        o << value << " ";
+    }
+    return o << "}";
+}


I'd prefer to have a pretty printer wrapper for this sort of thing, as the one formatter for all ranges is a bit of a Procrustean bed. Use would be nearly as simple though, e.g.

std::cout << range_printer(some_range) << ...

I see your point. That Procrustes sounds like quite an unsavoury character, so I am sorry if my one-size-fits-all helper function caused harm to any travellers crossing our code base.

halfflat · 2016-10-28T13:40:27Z

tests/unit/test_fvm_multi.cpp

+        for(auto val : v) if(val>=0.) return false;
+        return true;
+    };
+    EXPECT_TRUE(is_neg(J.l()(1, J.size())));


Formatting of test lambdas. (Consider using std::any_of/std::all_of?)

* Use divided compartments to determine FVM coefficients. * Pick correct control volume in FVM from sgement position (avoids off-by-half error.) * Add colour override functionality to tsplot: `--colour` option. * Add const accessor for cell soma. * Source formatting, comments in `math.hpp` * Fix `range_view`: was using incorrectly named type trait. * Add unit test for `range_view`. * Allow points of discontinuity to be omitted from L-infinity norm calculations. * Add `-d, --min-dt` option to `validate.exe` to control time step in validation convergence tests. * Add validation test: confirm divided compartment policy does not effect results on simple frustrum dendrites. * Change default max compartments on validation tests to 100 (ad hoc observed convergence limit at dt circa 0.001 ms; finder spatial division would required much finer dt.) * Make NEURON validation data generation scripts use CVODE by default, and with `secondorder=2` when non-zero `dt` is given.

* Use only `div_compartment_integrator` for compartmentalization in `fvm_multicell`. The policy will later be moved to a backend policy class. * For now, disable validation tests that test different division policies (see above). * Tweak comments and remove redundant `using`, following comments on PR#54.

Complex compartments

* first version of openmp threading back end * adding openmp parallel sort implementation * OpenMP sort working * Support for units syntax within state block. * Add soma-less cable cell to test cells. Also: * Ensure intrinsic and passive properties properly set on test cells. * Change bulk resistivity default. * Align defaults with values used in most of the NEURON validation scripts. * Use consistent 100 Ω·m bulk resistivity across both NEURON test models and basic validation cells. * OpenMP back end working * Add Extrae+paraver support, needs to fix compilation warnings * Reorganize validation data generation * Move generation and data to top-level validation directory. * Make BUILD_VALIDATION_DATA and VALIDATION_DATA_DIR cache vars. * Add helper CMake functions for data generation. Note `validation/ref/numeric/foo.sh` is just a placeholder. * Bugfix: hh_soma.jl * Use consistent scaling for y[1] scalar voltage in hh_soma.jl * Also: add more reserved target names to CMakeLists.txt helper function. * Refactor convergence tests; add numeric soma ref. * Amend data directory path in validation tests. * Enmodulate `hh_soma.jl` * Add HH channel reference data generations script. * Switch `validate_soma.cpp` to numeric reference data. * Consolidate common code in `validate_ball_and_stick.cpp` * Add (nearly) Rallpack1 validation test (see below). * Gentle failure on absence of reference data in `validate_ball_and_stick.cpp` Can't yet override mechanism default parameter values, so the cable cell model added to `test_common_cells.hpp` lets the default stand; validation script will have to use the default membrane conductance rather than that given by Rallpack1. * Add Rallpack1 validation, plus bugfix, clean * Implement Rallpack1 validation test (with a workaround for inability to set membrane conductance). * Fix bug in L≠1 case in PassiveCable.jl (this may still be wrong). * Fix bug in peak delta computation in trace analysis when both traces have no local maxima. * Gentle failure on missing `numeric_soma.json` * Allow multiple `-s` selection operations for `tsplot`, acting disjunctively. * Remove errant test file. * file's cleanup * Remove tabs * Use correct routine in numeric_rallpack1.jl x0.3 * Configure-time test for julia * `math::infinity<>()` wrapper for infinity * Use name `i_e` for Stim current density * Use `math::infinity<>()` for infinite value * Adds unit tests for the STATE block. * Add "lib" to search prefixes for libtbb * Fix quoting error in library search. * Add "lib" to prefixes when system is "Linux". * Address deprecated use of 'symbol' warning. Julia 0.5 deprecates use of `symbol` instead of `Symbol`. This patch just substitutes the correct call. * Address deprecated use of 'symbol' warning. Julia 0.5 deprecates use of `symbol` instead of `Symbol`. This patch just substitutes the correct call. * Addresses PR comments. * Unit tests for math.hpp * Tests for `math::pi`, `math::lerp`, `math::area_frustrum` and `math::volume_frustrum` * Fix `math:pi<long double>()`. * Extend range, view functionality. * New `filter` view: lazily selects based on predicate. * Generic `front` and `back` for sequences. * New rangeutil STL wrappers `stable_sort_by`, `all_of`, `any_of`. * Consolidate common utility unit testing structures into `tests/unit/common.hpp` * Add `ball_and_squiggle` model; fix `ball_and_taper`. * Make `test_common_cells.hpp` and `ball_and_taper.py` agree. * Add `ball_and_squiggle` model that has a tapering undulating profile. * Address PR#46 review comments. * Add documentation of template parameters for `filter_iterator`. * Document use of `uninitalized<F>` for holding functional objects in `filter_iterator` and `transform_iterator` * Consolidate validation test code (issue #41) * Simplify trace analysis and reporting code in `trace_analysis.hpp` * Consolidate convergence test run procedures into new class `convergence_test_runner`. * New compartment info structure for FVM. * Make `algorithm::sum`, `algorithm::mean` more generic, allowing use with array types. * Add `div_compartment` compartment representation, that holds geometric information for each half of a compartment that will then be used in calculating control volumes. * Add three compartmentalisation schemes/policies that discretize a segment into `div_compartment` objects: * `div_compartment_by_ends` divides based only on the segment end points and radii. * `div_compartment_sampler` forms frusta by sampling the segment radius at each compartment boundary * `div_compartment_integrator` computes the compartment areas and volumes exactly by summing all frustra in the intersection of the segment and the compartmnet span. * Extrae linked at execution time * cleaning project * Complex compartments * Use divided compartments to determine FVM coefficients. * Pick correct control volume in FVM from sgement position (avoids off-by-half error.) * Add colour override functionality to tsplot: `--colour` option. * Add const accessor for cell soma. * Source formatting, comments in `math.hpp` * Fix `range_view`: was using incorrectly named type trait. * Add unit test for `range_view`. * Allow points of discontinuity to be omitted from L-infinity norm calculations. * Add `-d, --min-dt` option to `validate.exe` to control time step in validation convergence tests. * Add validation test: confirm divided compartment policy does not effect results on simple frustrum dendrites. * Change default max compartments on validation tests to 100 (ad hoc observed convergence limit at dt circa 0.001 ms; finder spatial division would required much finer dt.) * Make NEURON validation data generation scripts use CVODE by default, and with `secondorder=2` when non-zero `dt` is given. * Remove division policy type parameter. * Use only `div_compartment_integrator` for compartmentalization in `fvm_multicell`. The policy will later be moved to a backend policy class. * For now, disable validation tests that test different division policies (see above). * Tweak comments and remove redundant `using`, following comments on PR#54. * Minor twicks and corrections

* streamline CMake * list of external libaries to link against CMake is concatenated in `EXTERNAL_LIBRARIES` variable * UNWIND_STACK is enabled by default if the libunwind package can be found * stack trace printer * optionally prints with color * takes the executable to perform address translation as command line argument * added documentation to c++, in particular some comments in fvm_multicell.hpp to explain some steps in more detail.

halfflat

Minor spleling fixes really, plus I'd like to see the libunwind.h include isolated.
Perhaps for another commit later, but I still believe that the indexed_view class should either be properly generic, or live in memory:: if it's tied to these storage classes.

halfflat · 2016-11-14T13:58:32Z

CMakeLists.txt

@@ -80,6 +85,7 @@ if(WITH_CUDA)

    add_definitions(-DWITH_GPU)
    include_directories(SYSTEM ${CUDA_INCLUDE_DIRS})
+    set(EXTERNAL_LIBRARIES ${EXTERNAL_LIBRARIES} ${CUDA_LIBRARIES})


I like this approach. Might be cleaner to use

list(APPEND EXTERNAL_LIBRARIES ${CUDA_LIBRARIES})

(and likewise above) instead of set.

halfflat · 2016-11-14T14:47:06Z

miniapp/CMakeLists.txt

@@ -18,21 +18,16 @@ else()
    add_executable(miniapp.exe ${MINIAPP_SOURCES} ${HEADERS})
 endif()

-target_link_libraries(miniapp.exe LINK_PUBLIC nestmc)
+set(aaa nestmc)


Is this debugging CMake code that snuck in?

halfflat · 2016-11-14T14:47:50Z

scripts/print_backtrace

@@ -23,27 +54,36 @@ def parse_backtrace(source):
            tokens = line.split()
            trace.append({'location':tokens[0], 'function':tokens[1]})
    else:
-        print "error: unable to open file ", source
+        print "error: unable to back trace file ", source


missing verb in error text?

halfflat · 2016-11-14T14:49:24Z

src/fvm_multicell.hpp

@@ -401,6 +400,15 @@ void fvm_multicell<Backend>::initialize(
    std::vector<value_type> tmp_cv_areas(ncomp);
    std::vector<value_type> tmp_cv_capacitance(ncomp);

+    // Iterate over the input cells and build the indexes etc that descrbe the
+    // fused cell group. On completion:
+    //  - group_paranet_index contains the full parent index for the fused cells.


spellling eror

note to self paranet is not a word

halfflat · 2016-11-14T14:51:44Z

src/util/unwind.cpp

@@ -51,8 +53,8 @@ std::string demangle(std::string s) {
    int status;
    char* demangled = abi::__cxa_demangle(s.c_str(), nullptr, nullptr, &status);

-    // __cxa_demangle only returns a valid string if the identifier in s was
-    // a mangled c++ symbol (i.e. returns an empty string for normal c symbols)
+    // the string returned by __cxa_demangle is on valid if it was passed a valid


mroe speeling

I don't like your tone.

halfflat · 2016-11-14T15:52:34Z

src/util/unwind.hpp

@@ -14,28 +9,31 @@ namespace mc {
 namespace util {

 #ifdef WITH_UNWIND


Given that the only reason libunwind.h needs to be included here is to get the definition of unw_word_t, which itself is an integral type, could we keep the #ifdefs and such inside unwind.cpp, hidden from the rest of the code, and use e.g.

struct source_location { std::string name; std::uintptr_t address; };

which does not depend upon the architecture-specific int size used for unw_word_t?

This has the nice side effect of keeping the global namespace free of the libunwind-specific enums and typedefs.

done. I was thinking about this as I was implementing this. I have assumed that it is a std::uint64_t, which is how it is defined in libunwind, with a static_assert in unwind.cpp, that will break the build if that assumption is every broken.

Any reason not to use uintptr_t?

bcumming and others added 30 commits September 8, 2016 16:10

nvcc can build miniapp

b38f78b

* add CMake option and target for CUDA * add some workarounds for CUDA compiler bugs

WIP towards basic GPU implementation

7e54957

* created a gpu multicell target type * missing CUDA specific kernels * basic cell state (voltage, current, etc) is on gpu * moved the vector library into main source tree * have not yet deleted external dependency

gpu WIP

b7bd5dd

Merge branch 'master' into gpu

3edad46

WIP: progress towards crash free GPU existance

1388d20

finish moving vector to internal memory module

8e6d97f

add the new tests that I forgot in the last commit

4ee9930

updated fvm multicell code for the new memory library

4c0864b

WIP gpu version

d1a5b6e

* more simplifications of the memory library * clean up of policy based matrix gpu/multicore implementation

WIP merge

36ebcc7

Merge branch 'master' into gpu

851f38c

gpu WIP

4a1777f

WIP gpu

c742fd5

Merge pull request #1 from eth-cscs/master

039d5a9

merge from main

WIP

121dade

Support for units syntax within state block.

b6aec69

WIP

44c8ac0

WIP

78f60f5

Merge branch 'master' of github.com:eth-cscs/nestmc-proto

5c09594

WIP

3830c9a

WIP

2db99a2

WIP

65ed2fe

WIP

7f1216d

merge with master

de992ad

WIP

b5bae5b

Merge branch 'master' of github.com:eth-cscs/nestmc-proto into gpu

06ef117

Merge branch 'master' into bugfix/modcc/state-block-units

18a9226

Adds unit tests for the STATE block.

05e4770

WIP

b9b1e57

bcumming mentioned this pull request Oct 28, 2016

More range functionality, unit tests. #46

Merged

halfflat and others added 7 commits October 28, 2016 11:09

Address PR#46 review comments.

cee495c

* Add documentation of template parameters for `filter_iterator`. * Document use of `uninitalized<F>` for holding functional objects in `filter_iterator` and `transform_iterator`

Merge pull request #46 from halfflat/feature/more-range-utils

c97135d

More range functionality, unit tests.

Merge pull request #47 from halfflat/feature/new-test-model

5ade8d0

Feature/new test model

Consolidate validation test code (issue #41)

1b929ff

* Simplify trace analysis and reporting code in `trace_analysis.hpp` * Consolidate convergence test run procedures into new class `convergence_test_runner`.

Merge pull request #48 from halfflat/feature/consolidate-validation-t…

550da10

…ests Consolidate validation tests (issue #41)

Merge pull request #49 from halfflat/feature/divided-compartments

a03af27

Feature/divided compartments

halfflat reviewed Oct 28, 2016

View reviewed changes

halfflat and others added 5 commits October 30, 2016 22:31

Merge pull request #54 from halfflat/feature/complex-compartments

e8d3285

Complex compartments

WIP

b39a93e

bcumming mentioned this pull request Nov 4, 2016

Clean up GPU branch [10] #61

Closed

bcumming added 5 commits November 4, 2016 20:41

WIP - gpu version runs but doesn't validate

cd4d9ae

WIP

e74720b

backend refactoring

b11990f

WIP

bb14d18

merge master into gpu branch (new validation tests mostly)

1584bf3

bcumming assigned halfflat Nov 11, 2016

bcumming added 3 commits November 11, 2016 17:31

rename template parameter in indexed_view to Backend

46427e6

findunwind works when libunwind is unavailable

2b2b310

halfflat suggested changes Nov 14, 2016

View reviewed changes

bcumming and others added 3 commits November 14, 2016 18:07

final clean up for PR #45

5825543

further simplifications to the unwind C++ code

faf1a41

Tiny typo fix

6cf11d3

halfflat approved these changes Nov 14, 2016

View reviewed changes

halfflat merged commit 1bc18ea into arbor-sim:gpu Nov 14, 2016

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Part 1 of the big GPU merge #45

Part 1 of the big GPU merge #45

bcumming commented Oct 27, 2016 •

edited

Loading

halfflat Oct 28, 2016

bcumming Nov 9, 2016

halfflat Oct 28, 2016

halfflat Oct 28, 2016

bcumming Nov 4, 2016

bcumming Nov 4, 2016

halfflat Oct 28, 2016

bcumming Nov 9, 2016

halfflat Oct 28, 2016

bcumming Nov 4, 2016

halfflat Nov 4, 2016

halfflat Oct 28, 2016

bcumming Nov 9, 2016

halfflat Oct 28, 2016

bcumming Nov 9, 2016

halfflat Nov 9, 2016

halfflat Oct 28, 2016

bcumming Nov 9, 2016

halfflat Oct 28, 2016

bcumming Oct 30, 2016

halfflat Oct 28, 2016

halfflat left a comment

halfflat Nov 14, 2016

halfflat Nov 14, 2016

halfflat Nov 14, 2016

halfflat Nov 14, 2016

bcumming Nov 14, 2016

halfflat Nov 14, 2016

bcumming Nov 14, 2016

halfflat Nov 14, 2016

bcumming Nov 14, 2016

halfflat Nov 14, 2016

		@@ -14,28 +9,31 @@ namespace mc {
		namespace util {

		#ifdef WITH_UNWIND

Part 1 of the big GPU merge #45

Part 1 of the big GPU merge #45

Conversation

bcumming commented Oct 27, 2016 • edited Loading

build System

modcc

algorithms

backends

lowered fvm cells

memory library

debug backtraces

util simplification and consolidation

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

halfflat left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bcumming commented Oct 27, 2016 •

edited

Loading