Merge pull request #22 from arghdos/spyJac-working

Fix sub-kernel wrapper generation & test

.travis.yml

@@ -13,22 +13,23 @@ language: python
 matrix:
   include:
     - python: "3.6"
-      env: ISFLAKE="0" PYJAC_TEST_LANGS="opencl"
+      env: PYJAC_TEST_LANGS="opencl"
     - python: "3.6"
-      env: ISFLAKE="0" PYJAC_TEST_LANGS="c"
+      env: PYJAC_TEST_LANGS="c"
     - python: "3.6"
-      env: ISFLAKE="0" PYJAC_TEST_LANGS="opencl" PYJAC_RXN_SORTING="simd"
+      env: PYJAC_TEST_LANGS="c" PYJAC_RXN_SORTING="simd"
     - python: "3.6"
-      env: ISFLAKE="0" PYJAC_TEST_LANGS="c" PYJAC_RXN_SORTING="simd"
+      env: PYJAC_TEST_LANGS="opencl" RUN_LONG_TESTS="1"
+    - python: "3.6"
+      env: PYJAC_TEST_LANGS="opencl" RUN_LONG_TESTS="1" PYJAC_UNIQUE_POINTERS="1"
+    - python: "3.6"
+      env: PYJAC_TEST_LANGS="c" RUN_LONG_TESTS="1"
+    - python: "3.6"
+      env: PYJAC_TEST_LANGS="c" RUN_LONG_TESTS="1" PYJAC_UNIQUE_POINTERS="1"
     - python: "3.6"
       env: ISFLAKE="1"
 
-# global env
-env:
-  global:
-    - PRIORITY=500
-    - LLV=5.0
-
+# additional packages
 addons:
   apt:
     sources:
@@ -66,7 +67,7 @@ before_install:
   - conda config --add channels conda-forge
   - conda config --add channels cantera
   # create env
-  - conda create -y -n test-environment python=$TRAVIS_PYTHON_VERSION llvmdev clangdev cantera ocl-icd islpy pyyaml scipy pyopencl numpy Cython pytables flake8 pep8-naming pocl adept
+  - conda create -y -n test-environment python=$TRAVIS_PYTHON_VERSION llvmdev clangdev cantera ocl-icd=*=h14c3975_1001 islpy pyyaml scipy pyopencl numpy Cython pytables flake8 pep8-naming pocl adept=*=he6fcbdd_3
   - source activate test-environment
   # get vendor dir
   - export OCL_ICD_VENDORS=$CONDA_PREFIX/etc/OpenCL/vendors
@@ -89,8 +90,10 @@ before_script:
   - export TMPDIR=$TRAVIS_BUILD_DIR/working/
 # run test
 script:
-  if [[ "$ISFLAKE" == "0" ]]; then
-      nosetests -a '!verylong' -s --with-timer;
+  if [[ -z "$ISFLAKE" && -z "$RUN_LONG_TESTS" ]]; then
+      nosetests -A 'not fullkernel and not verylong' -s --with-timer;
+  elif [[ -z "$ISFLAKE" && -n "$RUN_LONG_TESTS" ]]; then
+      nosetests -A 'fullkernel' -s --with-timer;
   else
       python -m flake8 pyjac/ --count;
   fi

MANIFEST.in

@@ -19,7 +19,7 @@ recursive-include pyjac/core *.py
 recursive-include pyjac/functional_tester *.py *.yaml
 recursive-include pyjac/loopy_utils *.py *.in
 recursive-include pyjac/libgen *.py
-recursive-include pyjac/kernel_utils *.py *.in *.h *.c *.oclh *.ocl
+recursive-include pyjac/kernel_utils *.py *.in *.hpp *.cpp *.oclh *.ocl
 recursive-include pyjac/performance_tester *.py
 recursive-include pyjac/pywrap *.py *.in
 recursive-include pyjac/tests *.py *.inp * .cti *.in *.yaml

docs/driver_functions.rst

@@ -56,9 +56,9 @@ enable conversion to/from the calling code's state variables to pyJac's state-ve
 
 .. _work-size:
 
-========================
-Specifying the Work-Size
-========================
+==========
+Work Sizes
+==========
 
 In pyJac, the work-size is defined as the total number of separate (potentially
 vectorized) evaluations of the chemical kinetic properties / source rates / Jacobian
@@ -123,6 +123,42 @@ For vectorized execution, the shape of the arrays changes slightly to
 where the `vector_width` is typically 2--8 for CPUs and MICs, and 64--1024 for GPUs
 (note: this corresponds to the block-size in CUDA).
 
+==========================
+Coupling to external codes
+==========================
+
+One downside of pyJac's data-storage format is that it requires each OpenMP thread
+/ OpenCL work-group to be passed the **addresses to the beginning of the arrays
+that contain memory for the entire set of threads / work-groups.**  In code this
+might look something like this:
+
+```
+    double* phi = (double*)malloc(work_size * (N_s + 1) * sizeof(double));
+    double* dphi = (double*)malloc(work_size * (N_s + 1) * sizeof(double));
+    double* rwk = (double*)malloc(work_size * kernel.requiredMemorySize());
+    // populate state vector
+    // copy temperatures
+    memcpy(&phi[0], &T[0], work_size * sizeof(double));
+    ...
+    // call source rates
+    species_rates(0, phi, dphi, rwk);
+```
+
+However, many codes operate on local-copies of the state vector array, e.g.:
+
+```
+    double phi[N_s];
+    double dphi[N_s];
+    double* rwk = (double*)malloc(kernel.requiredMemorySize());
+    //populate state vector
+    phi[0] = T;
+    ...
+    species_rates(0, phi, dphi, rwk);
+```
+
+For this sort of code, you may supply the `-up` or `--unique_pointers` flag to the
+command line interface of pyJac.
+
 
 =====================
 Lockstep-based driver

pyjac/core/array_creator.py

@@ -128,7 +128,7 @@ def _should_split(self, array):
 
         """
 
-        return len(array.shape) >= 2 or self.is_simd
+        return len(array.shape) >= 1 or self.is_simd
 
     def split_and_vec_axes(self, array):
         """
@@ -169,6 +169,19 @@ def split_and_vec_axes(self, array):
 
         return split_axis, vec_axis
 
+    def grow_axis(self, array):
+        """
+        Returns
+        -------
+        grow_axis: int
+            The integer index of the axis corresponding to the initial conditions,
+            see :ref:`vector_split`
+        """
+        if self._should_split(array):
+            # the grow axis is one IFF it's a F-ordered deep-split
+            return 1 if self.data_order == 'F' else 0
+        return 0
+
     def split_shape(self, array):
         """
         Returns the array shape that would result from splitting the supplied array
@@ -202,7 +215,7 @@ def split_shape(self, array):
         vector_width = None
         if self._should_split(array):
             # the grow axis is one IFF it's a F-ordered deep-split
-            grow_axis = 1 if self.data_order == 'F' else 0
+            grow_axis = self.grow_axis(array)
             split_axis, vec_axis = self.split_and_vec_axes(array)
             vector_width = self.depth if self.depth else self.width
             assert vector_width
@@ -612,7 +625,7 @@ def initial_condition_dimension_vars(loopy_opts, test_size, is_driver_kernel=Fal
         The initial condition dimension size variables
     """
 
-    if isinstance(test_size, int) or loopy_opts.work_size:
+    if isinstance(test_size, int) or loopy_opts.unique_pointers:
         return []
     if is_driver_kernel:
         return [work_size, problem_size]