Changes in / [36a2418:e5bbe64] in sasmodels

Location:

sasmodels

Files:

• kernel.py (modified) (2 diffs)
• kernelcl.py (modified) (35 diffs)
• kernelcuda.py (modified) (26 diffs)
• kerneldll.py (modified) (15 diffs)
• kernelpy.py (modified) (11 diffs)
• model_test.py (modified) (1 diff)
• models/rpa.c (modified) (6 diffs)

sasmodels/kernel.py

@@ -23 +23 @@
 # pylint: enable=unused-import
 
-
 class KernelModel(object):
     info = None  # type: ModelInfo
@@ -34 +33 @@
         # type: () -> None
         pass
-
 
 class Kernel(object):

sasmodels/kernelcl.py

@@ -61 +61 @@
 
 
-# Attempt to setup OpenCL. This may fail if the pyopencl package is not
+# Attempt to setup opencl. This may fail if the pyopencl package is not
 # installed or if it is installed but there are no devices available.
 try:
@@ -67 +67 @@
     from pyopencl import mem_flags as mf
     from pyopencl.characterize import get_fast_inaccurate_build_options
-    # Ask OpenCL for the default context so that we know that one exists.
+    # Ask OpenCL for the default context so that we know that one exists
     cl.create_some_context(interactive=False)
     HAVE_OPENCL = True
@@ -88 +88 @@
 # pylint: enable=unused-import
 
-
-# CRUFT: pyopencl < 2017.1 (as of June 2016 needs quotes around include path).
+# CRUFT: pyopencl < 2017.1  (as of June 2016 needs quotes around include path)
 def quote_path(v):
     """
@@ -100 +99 @@
     return '"'+v+'"' if v and ' ' in v and not v[0] in "\"'-" else v
 
-
 def fix_pyopencl_include():
     """
@@ -107 +105 @@
     import pyopencl as cl
     if hasattr(cl, '_DEFAULT_INCLUDE_OPTIONS'):
-        cl._DEFAULT_INCLUDE_OPTIONS = [
-            quote_path(v) for v in cl._DEFAULT_INCLUDE_OPTIONS
-            ]
-
+        cl._DEFAULT_INCLUDE_OPTIONS = [quote_path(v) for v in cl._DEFAULT_INCLUDE_OPTIONS]
 
 if HAVE_OPENCL:
@@ -123 +118 @@
 MAX_LOOPS = 2048
 
+
 # Pragmas for enable OpenCL features.  Be sure to protect them so that they
 # still compile even if OpenCL is not present.
@@ -137 +133 @@
 """
 
-
 def use_opencl():
     sas_opencl = os.environ.get("SAS_OPENCL", "OpenCL").lower()
     return HAVE_OPENCL and sas_opencl != "none" and not sas_opencl.startswith("cuda")
 
-
 ENV = None
 def reset_environment():
@@ -150 +144 @@
     global ENV
     ENV = GpuEnvironment() if use_opencl() else None
-
 
 def environment():
@@ -168 +161 @@
     return ENV
 
-
 def has_type(device, dtype):
     # type: (cl.Device, np.dtype) -> bool
@@ -179 +171 @@
         return "cl_khr_fp64" in device.extensions
     else:
-        # Not supporting F16 type since it isn't accurate enough.
+        # Not supporting F16 type since it isn't accurate enough
         return False
-
 
 def get_warp(kernel, queue):
@@ -191 +182 @@
         cl.kernel_work_group_info.PREFERRED_WORK_GROUP_SIZE_MULTIPLE,
         queue.device)
-
 
 def compile_model(context, source, dtype, fast=False):
@@ -213 +203 @@
         source_list.insert(0, _F64_PRAGMA)
 
-    # Note: USE_SINCOS makes the Intel CPU slower under OpenCL.
+    # Note: USE_SINCOS makes the intel cpu slower under opencl
     if context.devices[0].type == cl.device_type.GPU:
         source_list.insert(0, "#define USE_SINCOS\n")
@@ -220 +210 @@
     source = "\n".join(source_list)
     program = cl.Program(context, source).build(options=options)
-
     #print("done with "+program)
     return program
 
 
-# For now, this returns one device in the context.
-# TODO: Create a context that contains all devices on all platforms.
+# for now, this returns one device in the context
+# TODO: create a context that contains all devices on all platforms
 class GpuEnvironment(object):
     """
-    GPU context for OpenCL, with possibly many devices and one queue per device.
+    GPU context, with possibly many devices, and one queue per device.
+
+    Because the environment can be reset during a live program (e.g., if the
+    user changes the active GPU device in the GUI), everything associated
+    with the device context must be cached in the environment and recreated
+    if the environment changes.  The *cache* attribute is a simple dictionary
+    which holds keys and references to objects, such as compiled kernels and
+    allocated buffers.  The running program should check in the cache for
+    long lived objects and create them if they are not there.  The program
+    should not hold onto cached objects, but instead only keep them active
+    for the duration of a function call.  When the environment is destroyed
+    then the *release* method for each active cache item is called before
+    the environment is freed.  This means that each cl buffer should be
+    in its own cache entry.
     """
     def __init__(self):
         # type: () -> None
-        # Find gpu context.
+        # find gpu context
         context_list = _create_some_context()
 
@@ -247 +249 @@
                 self.context[dtype] = None
 
-        # Build a queue for each context.
+        # Build a queue for each context
         self.queue = {}
         context = self.context[F32]
@@ -257 +259 @@
             self.queue[F64] = cl.CommandQueue(context, context.devices[0])
 
-        ## Byte boundary for data alignment.
+        # Byte boundary for data alignment
         #self.data_boundary = max(context.devices[0].min_data_type_align_size
         #                         for context in self.context.values())
 
-        # Cache for compiled programs, and for items in context.
+        # Cache for compiled programs, and for items in context
         self.compiled = {}
+        self.cache = {}
 
     def has_type(self, dtype):
@@ -277 +280 @@
         """
         # Note: PyOpenCL caches based on md5 hash of source, options and device
-        # but I'll do so as well just to save some data munging time.
+        # so we don't really need to cache things for ourselves.  I'll do so
+        # anyway just to save some data munging time.
         tag = generate.tag_source(source)
         key = "%s-%s-%s%s"%(name, dtype, tag, ("-fast" if fast else ""))
-        # Check timestamp on program.
+        # Check timestamp on program
         program, program_timestamp = self.compiled.get(key, (None, np.inf))
         if program_timestamp < timestamp:
@@ -293 +297 @@
         return program
 
+    def free_buffer(self, key):
+        if key in self.cache:
+            self.cache[key].release()
+            del self.cache[key]
+
+    def __del__(self):
+        for v in self.cache.values():
+            release = getattr(v, 'release', lambda: None)
+            release()
+        self.cache = {}
+
+_CURRENT_ID = 0
+def unique_id():
+    global _CURRENT_ID
+    _CURRENT_ID += 1
+    return _CURRENT_ID
 
 def _create_some_context():
@@ -305 +325 @@
     which one (and not a CUDA device, or no GPU).
     """
-    # Assume we do not get here if SAS_OPENCL is None or CUDA.
+    # Assume we do not get here if SAS_OPENCL is None or CUDA
     sas_opencl = os.environ.get('SAS_OPENCL', 'opencl')
     if sas_opencl.lower() != 'opencl':
-        # Setting PYOPENCL_CTX as a SAS_OPENCL to create cl context.
+        # Setting PYOPENCL_CTX as a SAS_OPENCL to create cl context
         os.environ["PYOPENCL_CTX"] = sas_opencl
 
@@ -316 +336 @@
         except Exception as exc:
             warnings.warn(str(exc))
-            warnings.warn("pyopencl.create_some_context() failed.  The "
-                "environment variable 'SAS_OPENCL' or 'PYOPENCL_CTX' might "
-                "not be set correctly")
+            warnings.warn("pyopencl.create_some_context() failed")
+            warnings.warn("the environment variable 'SAS_OPENCL' or 'PYOPENCL_CTX' might not be set correctly")
 
     return _get_default_context()
-
 
 def _get_default_context():
@@ -334 +352 @@
     # is running may increase throughput.
     #
-    # MacBook Pro, base install:
+    # Macbook pro, base install:
     #     {'Apple': [Intel CPU, NVIDIA GPU]}
-    # MacBook Pro, base install:
+    # Macbook pro, base install:
     #     {'Apple': [Intel CPU, Intel GPU]}
-    # 2 x NVIDIA 295 with Intel and NVIDIA opencl drivers install:
+    # 2 x nvidia 295 with Intel and NVIDIA opencl drivers installed
     #     {'Intel': [CPU], 'NVIDIA': [GPU, GPU, GPU, GPU]}
     gpu, cpu = None, None
@@ -361 +379 @@
             else:
                 # System has cl.device_type.ACCELERATOR or cl.device_type.CUSTOM
-                # Intel Phi for example registers as an accelerator.
+                # Intel Phi for example registers as an accelerator
                 # Since the user installed a custom device on their system
                 # and went through the pain of sorting out OpenCL drivers for
@@ -368 +386 @@
                 gpu = device
 
-    # Order the devices by gpu then by cpu; when searching for an available
+    # order the devices by gpu then by cpu; when searching for an available
     # device by data type they will be checked in this order, which means
     # that if the gpu supports double then the cpu will never be used (though
@@ -395 +413 @@
     that the compiler is allowed to take shortcuts.
     """
-    info = None  # type: ModelInfo
-    source = ""  # type: str
-    dtype = None  # type: np.dtype
-    fast = False  # type: bool
-    _program = None  # type: cl.Program
-    _kernels = None  # type: Dict[str, cl.Kernel]
-
     def __init__(self, source, model_info, dtype=generate.F32, fast=False):
         # type: (Dict[str,str], ModelInfo, np.dtype, bool) -> None
@@ -408 +419 @@
         self.dtype = dtype
         self.fast = fast
+        self.timestamp = generate.ocl_timestamp(self.info)
+        self._cache_key = unique_id()
 
     def __getstate__(self):
@@ -416 +429 @@
         # type: (Tuple[ModelInfo, str, np.dtype, bool]) -> None
         self.info, self.source, self.dtype, self.fast = state
-        self._program = self._kernels = None
 
     def make_kernel(self, q_vectors):
@@ -422 +434 @@
         return GpuKernel(self, q_vectors)
 
-    def get_function(self, name):
+    @property
+    def Iq(self):
+        return self._fetch_kernel('Iq')
+
+    def fetch_kernel(self, name):
         # type: (str) -> cl.Kernel
         """
@@ -428 +444 @@
         does not already exist.
         """
-        if self._program is None:
-            self._prepare_program()
-        return self._kernels[name]
-
-    def _prepare_program(self):
-        # type: (str) -> None
-        env = environment()
-        timestamp = generate.ocl_timestamp(self.info)
-        program = env.compile_program(
-            self.info.name,
-            self.source['opencl'],
-            self.dtype,
-            self.fast,
-            timestamp)
-        variants = ['Iq', 'Iqxy', 'Imagnetic']
-        names = [generate.kernel_name(self.info, k) for k in variants]
-        functions = [getattr(program, k) for k in names]
-        self._kernels = {k: v for k, v in zip(variants, functions)}
-        # Keep a handle to program so GC doesn't collect.
-        self._program = program
-
-
-# TODO: Check that we don't need a destructor for buffers which go out of scope.
+        gpu = environment()
+        key = self._cache_key
+        if key not in gpu.cache:
+            program = gpu.compile_program(
+                self.info.name,
+                self.source['opencl'],
+                self.dtype,
+                self.fast,
+                self.timestamp)
+            variants = ['Iq', 'Iqxy', 'Imagnetic']
+            names = [generate.kernel_name(self.info, k) for k in variants]
+            kernels = [getattr(program, k) for k in names]
+            data = dict((k, v) for k, v in zip(variants, kernels))
+            # keep a handle to program so GC doesn't collect
+            data['program'] = program
+            gpu.cache[key] = data
+        else:
+            data = gpu.cache[key]
+        return data[name]
+
+# TODO: check that we don't need a destructor for buffers which go out of scope
 class GpuInput(object):
     """
@@ -472 +486 @@
     def __init__(self, q_vectors, dtype=generate.F32):
         # type: (List[np.ndarray], np.dtype) -> None
-        # TODO: Do we ever need double precision q?
+        # TODO: do we ever need double precision q?
         self.nq = q_vectors[0].size
         self.dtype = np.dtype(dtype)
         self.is_2d = (len(q_vectors) == 2)
-        # TODO: Stretch input based on get_warp().
-        # Not doing it now since warp depends on kernel, which is not known
+        # TODO: stretch input based on get_warp()
+        # not doing it now since warp depends on kernel, which is not known
         # at this point, so instead using 32, which is good on the set of
         # architectures tested so far.
@@ -490 +504 @@
             self.q[:self.nq] = q_vectors[0]
         self.global_size = [self.q.shape[0]]
-        #print("creating inputs of size", self.global_size)
-
-        # Transfer input value to GPU.
+        self._cache_key = unique_id()
+
+    @property
+    def q_b(self):
+        """Lazy creation of q buffer so it can survive context reset"""
         env = environment()
-        context = env.context[self.dtype]
-        self.q_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
-                             hostbuf=self.q)
+        key = self._cache_key
+        if key not in env.cache:
+            context = env.context[self.dtype]
+            #print("creating inputs of size", self.global_size)
+            buffer = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
+                               hostbuf=self.q)
+            env.cache[key] = buffer
+        return env.cache[key]
 
     def release(self):
         # type: () -> None
         """
-        Free the buffer associated with the q value.
-        """
-        if self.q_b is not None:
-            self.q_b.release()
-            self.q_b = None
+        Free the buffer associated with the q value
+        """
+        environment().free_buffer(id(self))
 
     def __del__(self):
@@ -511 +530 @@
         self.release()
 
-
 class GpuKernel(Kernel):
     """
@@ -518 +536 @@
     *model* is the GpuModel object to call
 
-    The kernel is derived from :class:`Kernel`, providing the
-    :meth:`call_kernel` method to evaluate the kernel for a given set of
-    parameters.  Because of the need to move the q values to the GPU before
-    evaluation, the kernel is instantiated for a particular set of q vectors,
-    and can be called many times without transfering q each time.
+    The following attributes are defined:
+
+    *info* is the module information
+
+    *dtype* is the kernel precision
+
+    *dim* is '1d' or '2d'
+
+    *result* is a vector to contain the results of the call
+
+    The resulting call method takes the *pars*, a list of values for
+    the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight)
+    vectors for the polydisperse parameters.  *cutoff* determines the
+    integration limits: any points with combined weight less than *cutoff*
+    will not be calculated.
 
     Call :meth:`release` when done with the kernel instance.
     """
-    #: SAS model information structure.
-    info = None  # type: ModelInfo
-    #: Kernel precision.
-    dtype = None  # type: np.dtype
-    #: Kernel dimensions (1d or 2d).
-    dim = ""  # type: str
-    #: Calculation results, updated after each call to :meth:`_call_kernel`.
-    result = None  # type: np.ndarray
-
     def __init__(self, model, q_vectors):
-        # type: (GpuModel, List[np.ndarray]) -> None
+        # type: (cl.Kernel, np.dtype, ModelInfo, List[np.ndarray]) -> None
         dtype = model.dtype
         self.q_input = GpuInput(q_vectors, dtype)
         self._model = model
-
-        # Attributes accessed from the outside.
+        # F16 isn't sufficient, so don't support it
+        self._as_dtype = np.float64 if dtype == generate.F64 else np.float32
+        self._cache_key = unique_id()
+
+        # attributes accessed from the outside
         self.dim = '2d' if self.q_input.is_2d else '1d'
         self.info = model.info
-        self.dtype = dtype
-
-        # Converter to translate input to target type.
-        self._as_dtype = np.float64 if dtype == generate.F64 else np.float32
-
-        # Holding place for the returned value.
+        self.dtype = model.dtype
+
+        # holding place for the returned value
         nout = 2 if self.info.have_Fq and self.dim == '1d' else 1
-        extra_q = 4  # Total weight, form volume, shell volume and R_eff.
-        self.result = np.empty(self.q_input.nq*nout + extra_q, dtype)
-
-        # Allocate result value on GPU.
+        extra_q = 4  # total weight, form volume, shell volume and R_eff
+        self.result = np.empty(self.q_input.nq*nout+extra_q, dtype)
+
+    @property
+    def _result_b(self):
+        """Lazy creation of result buffer so it can survive context reset"""
         env = environment()
-        context = env.context[self.dtype]
-        width = ((self.result.size+31)//32)*32 * self.dtype.itemsize
-        self._result_b = cl.Buffer(context, mf.READ_WRITE, width)
-
-    def _call_kernel(self, call_details, values, cutoff, magnetic,
-                     effective_radius_type):
-        # type: (CallDetails, np.ndarray, float, bool, int) -> np.ndarray
+        key = self._cache_key
+        if key not in env.cache:
+            context = env.context[self.dtype]
+            width = ((self.result.size+31)//32)*32 * self.dtype.itemsize
+            buffer = cl.Buffer(context, mf.READ_WRITE, width)
+            env.cache[key] = buffer
+        return env.cache[key]
+
+    def _call_kernel(self, call_details, values, cutoff, magnetic, effective_radius_type):
+        # type: (CallDetails, np.ndarray, np.ndarray, float, bool) -> np.ndarray
         env = environment()
         queue = env.queue[self._model.dtype]
         context = queue.context
 
-        # Arrange data transfer to card.
+        # Arrange data transfer to/from card
+        q_b = self.q_input.q_b
+        result_b = self._result_b
         details_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
                               hostbuf=call_details.buffer)
@@ -573 +599 @@
                              hostbuf=values)
 
-        # Setup kernel function and arguments.
         name = 'Iq' if self.dim == '1d' else 'Imagnetic' if magnetic else 'Iqxy'
-        kernel = self._model.get_function(name)
+        kernel = self._model.fetch_kernel(name)
         kernel_args = [
-            np.uint32(self.q_input.nq),  # Number of inputs.
-            None,  # Placeholder for pd_start.
-            None,  # Placeholder for pd_stop.
-            details_b,  # Problem definition.
-            values_b,  # Parameter values.
-            self.q_input.q_b,  # Q values.
-            self._result_b,   # Result storage.
-            self._as_dtype(cutoff),  # Probability cutoff.
-            np.uint32(effective_radius_type),  # R_eff mode.
+            np.uint32(self.q_input.nq), None, None,
+            details_b, values_b, q_b, result_b,
+            self._as_dtype(cutoff),
+            np.uint32(effective_radius_type),
         ]
-
-        # Call kernel and retrieve results.
         #print("Calling OpenCL")
         #call_details.show(values)
+        #Call kernel and retrieve results
         wait_for = None
         last_nap = time.clock()
@@ -601 +620 @@
                                *kernel_args, wait_for=wait_for)]
             if stop < call_details.num_eval:
-                # Allow other processes to run.
+                # Allow other processes to run
                 wait_for[0].wait()
                 current_time = time.clock()
@@ -607 +626 @@
                     time.sleep(0.001)
                     last_nap = current_time
-        cl.enqueue_copy(queue, self.result, self._result_b, wait_for=wait_for)
+        cl.enqueue_copy(queue, self.result, result_b, wait_for=wait_for)
         #print("result", self.result)
 
-        # Free buffers.
-        details_b.release()
-        values_b.release()
+        # Free buffers
+        for v in (details_b, values_b):
+            if v is not None:
+                v.release()
 
     def release(self):
@@ -619 +639 @@
         Release resources associated with the kernel.
         """
+        environment().free_buffer(id(self))
         self.q_input.release()
-        if self._result_b is not None:
-            self._result_b.release()
-            self._result_b = None
 
     def __del__(self):

sasmodels/kernelcuda.py

@@ -63 +63 @@
 import time
 import re
-import atexit
 
 import numpy as np  # type: ignore
 
 
-# Attempt to setup CUDA. This may fail if the pycuda package is not
+# Attempt to setup cuda. This may fail if the pycuda package is not
 # installed or if it is installed but there are no devices available.
 try:
@@ -108 +107 @@
 MAX_LOOPS = 2048
 
-
 def use_cuda():
-    sas_opencl = os.environ.get("SAS_OPENCL", "CUDA").lower()
-    return HAVE_CUDA and sas_opencl.startswith("cuda")
-
+    env = os.environ.get("SAS_OPENCL", "").lower()
+    return HAVE_CUDA and (env == "" or env.startswith("cuda"))
 
 ENV = None
@@ -124 +121 @@
         ENV.release()
     ENV = GpuEnvironment() if use_cuda() else None
-
 
 def environment():
@@ -142 +138 @@
     return ENV
 
-
-# PyTest is not freeing ENV, so make sure it gets freed.
-atexit.register(lambda: ENV.release() if ENV is not None else None)
-
-
 def has_type(dtype):
     # type: (np.dtype) -> bool
@@ -152 +143 @@
     Return true if device supports the requested precision.
     """
-    # Assume the NVIDIA card supports 32-bit and 64-bit floats.
-    # TODO: Check if pycuda support F16.
+    # Assume the nvidia card supports 32-bit and 64-bit floats.
+    # TODO: check if pycuda support F16
     return dtype in (generate.F32, generate.F64)
 
 
 FUNCTION_PATTERN = re.compile(r"""^
-  (?P<space>\s*)                       # Initial space.
-  (?P<qualifiers>^(?:\s*\b\w+\b\s*)+)  # One or more qualifiers before function.
-  (?P<function>\s*\b\w+\b\s*[(])       # Function name plus open parens.
+  (?P<space>\s*)                   # initial space
+  (?P<qualifiers>^(?:\s*\b\w+\b\s*)+) # one or more qualifiers before function
+  (?P<function>\s*\b\w+\b\s*[(])      # function name plus open parens
   """, re.VERBOSE|re.MULTILINE)
 
@@ -167 +158 @@
   """, re.VERBOSE|re.MULTILINE)
 
-
 def _add_device_tag(match):
     # type: (None) -> str
-    # Note: Should be re.Match, but that isn't a simple type.
+    # Note: should be re.Match, but that isn't a simple type
     """
     replace qualifiers with __device__ qualifiers if needed
@@ -183 +173 @@
         return "".join((space, "__device__ ", qualifiers, function))
 
-
 def mark_device_functions(source):
     # type: (str) -> str
@@ -190 +179 @@
     """
     return FUNCTION_PATTERN.sub(_add_device_tag, source)
-
 
 def show_device_functions(source):
@@ -200 +188 @@
         print(match.group('qualifiers').replace('\n',r'\n'), match.group('function'), '(')
     return source
-
 
 def compile_model(source, dtype, fast=False):
@@ -225 +212 @@
     #options = ['--verbose', '-E']
     options = ['--use_fast_math'] if fast else None
-    program = SourceModule(source, no_extern_c=True, options=options) #, include_dirs=[...])
+    program = SourceModule(source, no_extern_c=True, options=options) # include_dirs=[...]
 
     #print("done with "+program)
@@ -231 +218 @@
 
 
-# For now, this returns one device in the context.
-# TODO: Create a context that contains all devices on all platforms.
+# for now, this returns one device in the context
+# TODO: create a context that contains all devices on all platforms
 class GpuEnvironment(object):
     """
-    GPU context for CUDA.
+    GPU context, with possibly many devices, and one queue per device.
     """
     context = None # type: cuda.Context
     def __init__(self, devnum=None):
         # type: (int) -> None
+        # Byte boundary for data alignment
+        #self.data_boundary = max(d.min_data_type_align_size
+        #                         for d in self.context.devices)
+        self.compiled = {}
         env = os.environ.get("SAS_OPENCL", "").lower()
         if devnum is None and env.startswith("cuda:"):
             devnum = int(env[5:])
-
         # Set the global context to the particular device number if one is
         # given, otherwise use the default context.  Perhaps this will be set
@@ -252 +242 @@
             self.context = make_default_context()
 
-        ## Byte boundary for data alignment.
-        #self.data_boundary = max(d.min_data_type_align_size
-        #                         for d in self.context.devices)
-
-        # Cache for compiled programs, and for items in context.
-        self.compiled = {}
-
     def release(self):
         if self.context is not None:
@@ -279 +262 @@
         Compile the program for the device in the given context.
         """
-        # Note: PyCuda (probably) caches but I'll do so as well just to
-        # save some data munging time.
+        # Note: PyOpenCL caches based on md5 hash of source, options and device
+        # so we don't really need to cache things for ourselves.  I'll do so
+        # anyway just to save some data munging time.
         tag = generate.tag_source(source)
         key = "%s-%s-%s%s"%(name, dtype, tag, ("-fast" if fast else ""))
-        # Check timestamp on program.
+        # Check timestamp on program
         program, program_timestamp = self.compiled.get(key, (None, np.inf))
         if program_timestamp < timestamp:
@@ -293 +277 @@
         return program
 
-
 class GpuModel(KernelModel):
     """
@@ -309 +292 @@
     that the compiler is allowed to take shortcuts.
     """
-    info = None  # type: ModelInfo
-    source = ""  # type: str
-    dtype = None  # type: np.dtype
-    fast = False  # type: bool
-    _program = None  # type: SourceModule
-    _kernels = None  # type: Dict[str, cuda.Function]
+    info = None # type: ModelInfo
+    source = "" # type: str
+    dtype = None # type: np.dtype
+    fast = False # type: bool
+    program = None # type: SourceModule
+    _kernels = None # type: List[cuda.Function]
 
     def __init__(self, source, model_info, dtype=generate.F32, fast=False):
@@ -322 +305 @@
         self.dtype = dtype
         self.fast = fast
+        self.program = None # delay program creation
+        self._kernels = None
 
     def __getstate__(self):
@@ -330 +315 @@
         # type: (Tuple[ModelInfo, str, np.dtype, bool]) -> None
         self.info, self.source, self.dtype, self.fast = state
-        self._program = self._kernels = None
+        self.program = None
 
     def make_kernel(self, q_vectors):
         # type: (List[np.ndarray]) -> "GpuKernel"
-        return GpuKernel(self, q_vectors)
-
-    def get_function(self, name):
-        # type: (str) -> cuda.Function
-        """
-        Fetch the kernel from the environment by name, compiling it if it
-        does not already exist.
-        """
-        if self._program is None:
-            self._prepare_program()
-        return self._kernels[name]
-
-    def _prepare_program(self):
-        # type: (str) -> None
-        env = environment()
-        timestamp = generate.ocl_timestamp(self.info)
-        program = env.compile_program(
-            self.info.name,
-            self.source['opencl'],
-            self.dtype,
-            self.fast,
-            timestamp)
-        variants = ['Iq', 'Iqxy', 'Imagnetic']
-        names = [generate.kernel_name(self.info, k) for k in variants]
-        functions = [program.get_function(k) for k in names]
-        self._kernels = {k: v for k, v in zip(variants, functions)}
-        # Keep a handle to program so GC doesn't collect.
-        self._program = program
-
-
-# TODO: Check that we don't need a destructor for buffers which go out of scope.
+        if self.program is None:
+            compile_program = environment().compile_program
+            timestamp = generate.ocl_timestamp(self.info)
+            self.program = compile_program(
+                self.info.name,
+                self.source['opencl'],
+                self.dtype,
+                self.fast,
+                timestamp)
+            variants = ['Iq', 'Iqxy', 'Imagnetic']
+            names = [generate.kernel_name(self.info, k) for k in variants]
+            kernels = [self.program.get_function(k) for k in names]
+            self._kernels = dict((k, v) for k, v in zip(variants, kernels))
+        is_2d = len(q_vectors) == 2
+        if is_2d:
+            kernel = [self._kernels['Iqxy'], self._kernels['Imagnetic']]
+        else:
+            kernel = [self._kernels['Iq']]*2
+        return GpuKernel(kernel, self.dtype, self.info, q_vectors)
+
+    def release(self):
+        # type: () -> None
+        """
+        Free the resources associated with the model.
+        """
+        if self.program is not None:
+            self.program = None
+
+    def __del__(self):
+        # type: () -> None
+        self.release()
+
+# TODO: check that we don't need a destructor for buffers which go out of scope
 class GpuInput(object):
     """
@@ -386 +373 @@
     def __init__(self, q_vectors, dtype=generate.F32):
         # type: (List[np.ndarray], np.dtype) -> None
-        # TODO: Do we ever need double precision q?
+        # TODO: do we ever need double precision q?
         self.nq = q_vectors[0].size
         self.dtype = np.dtype(dtype)
         self.is_2d = (len(q_vectors) == 2)
-        # TODO: Stretch input based on get_warp().
-        # Not doing it now since warp depends on kernel, which is not known
+        # TODO: stretch input based on get_warp()
+        # not doing it now since warp depends on kernel, which is not known
         # at this point, so instead using 32, which is good on the set of
         # architectures tested so far.
         if self.is_2d:
-            width = ((self.nq+15)//16)*16
+            # Note: 16 rather than 15 because result is 1 longer than input.
+            width = ((self.nq+16)//16)*16
             self.q = np.empty((width, 2), dtype=dtype)
             self.q[:self.nq, 0] = q_vectors[0]
             self.q[:self.nq, 1] = q_vectors[1]
         else:
-            width = ((self.nq+31)//32)*32
+            # Note: 32 rather than 31 because result is 1 longer than input.
+            width = ((self.nq+32)//32)*32
             self.q = np.empty(width, dtype=dtype)
             self.q[:self.nq] = q_vectors[0]
         self.global_size = [self.q.shape[0]]
         #print("creating inputs of size", self.global_size)
-
-        # Transfer input value to GPU.
         self.q_b = cuda.to_device(self.q)
 
@@ -412 +399 @@
         # type: () -> None
         """
-        Free the buffer associated with the q value.
+        Free the memory.
         """
         if self.q_b is not None:
@@ -422 +409 @@
         self.release()
 
-
 class GpuKernel(Kernel):
     """
     Callable SAS kernel.
 
-    *model* is the GpuModel object to call
-
-    The kernel is derived from :class:`Kernel`, providing the
-    :meth:`call_kernel` method to evaluate the kernel for a given set of
-    parameters.  Because of the need to move the q values to the GPU before
-    evaluation, the kernel is instantiated for a particular set of q vectors,
-    and can be called many times without transfering q each time.
+    *kernel* is the GpuKernel object to call
+
+    *model_info* is the module information
+
+    *q_vectors* is the q vectors at which the kernel should be evaluated
+
+    *dtype* is the kernel precision
+
+    The resulting call method takes the *pars*, a list of values for
+    the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight)
+    vectors for the polydisperse parameters.  *cutoff* determines the
+    integration limits: any points with combined weight less than *cutoff*
+    will not be calculated.
 
     Call :meth:`release` when done with the kernel instance.
     """
-    #: SAS model information structure.
-    info = None  # type: ModelInfo
-    #: Kernel precision.
-    dtype = None  # type: np.dtype
-    #: Kernel dimensions (1d or 2d).
-    dim = ""  # type: str
-    #: Calculation results, updated after each call to :meth:`_call_kernel`.
-    result = None  # type: np.ndarray
-
-    def __init__(self, model, q_vectors):
-        # type: (GpuModel, List[np.ndarray]) -> None
-        dtype = model.dtype
+    def __init__(self, kernel, dtype, model_info, q_vectors):
+        # type: (cl.Kernel, np.dtype, ModelInfo, List[np.ndarray]) -> None
         self.q_input = GpuInput(q_vectors, dtype)
-        self._model = model
-
-        # Attributes accessed from the outside.
+        self.kernel = kernel
+        # F16 isn't sufficient, so don't support it
+        self._as_dtype = np.float64 if dtype == generate.F64 else np.float32
+
+        # attributes accessed from the outside
         self.dim = '2d' if self.q_input.is_2d else '1d'
-        self.info = model.info
+        self.info = model_info
         self.dtype = dtype
 
-        # Converter to translate input to target type.
-        self._as_dtype = np.float64 if dtype == generate.F64 else np.float32
-
-        # Holding place for the returned value.
+        # holding place for the returned value
         nout = 2 if self.info.have_Fq and self.dim == '1d' else 1
-        extra_q = 4  # Total weight, form volume, shell volume and R_eff.
-        self.result = np.empty(self.q_input.nq*nout + extra_q, dtype)
-
-        # Allocate result value on GPU.
+        extra_q = 4  # total weight, form volume, shell volume and R_eff
+        self.result = np.empty(self.q_input.nq*nout+extra_q, dtype)
+
+        # Inputs and outputs for each kernel call
+        # Note: res may be shorter than res_b if global_size != nq
         width = ((self.result.size+31)//32)*32 * self.dtype.itemsize
-        self._result_b = cuda.mem_alloc(width)
-
-    def _call_kernel(self, call_details, values, cutoff, magnetic,
-                     effective_radius_type):
-        # type: (CallDetails, np.ndarray, float, bool, int) -> np.ndarray
-
-        # Arrange data transfer to card.
+        self.result_b = cuda.mem_alloc(width)
+        self._need_release = [self.result_b]
+
+    def _call_kernel(self, call_details, values, cutoff, magnetic, effective_radius_type):
+        # type: (CallDetails, np.ndarray, np.ndarray, float, bool) -> np.ndarray
+        # Arrange data transfer to card
         details_b = cuda.to_device(call_details.buffer)
         values_b = cuda.to_device(values)
 
-        # Setup kernel function and arguments.
-        name = 'Iq' if self.dim == '1d' else 'Imagnetic' if magnetic else 'Iqxy'
-        kernel = self._model.get_function(name)
-        kernel_args = [
-            np.uint32(self.q_input.nq),  # Number of inputs.
-            None,  # Placeholder for pd_start.
-            None,  # Placeholder for pd_stop.
-            details_b,  # Problem definition.
-            values_b,  # Parameter values.
-            self.q_input.q_b,  # Q values.
-            self._result_b,   # Result storage.
-            self._as_dtype(cutoff),  # Probability cutoff.
-            np.uint32(effective_radius_type),  # R_eff mode.
+        kernel = self.kernel[1 if magnetic else 0]
+        args = [
+            np.uint32(self.q_input.nq), None, None,
+            details_b, values_b, self.q_input.q_b, self.result_b,
+            self._as_dtype(cutoff),
+            np.uint32(effective_radius_type),
         ]
         grid = partition(self.q_input.nq)
-
-        # Call kernel and retrieve results.
-        #print("Calling CUDA")
+        #print("Calling OpenCL")
         #call_details.show(values)
+        # Call kernel and retrieve results
         last_nap = time.clock()
         step = 100000000//self.q_input.nq + 1
@@ -502 +475 @@
             stop = min(start + step, call_details.num_eval)
             #print("queuing",start,stop)
-            kernel_args[1:3] = [np.int32(start), np.int32(stop)]
-            kernel(*kernel_args, **grid)
+            args[1:3] = [np.int32(start), np.int32(stop)]
+            kernel(*args, **grid)
             if stop < call_details.num_eval:
                 sync()
-                # Allow other processes to run.
+                # Allow other processes to run
                 current_time = time.clock()
                 if current_time - last_nap > 0.5:
@@ -512 +485 @@
                     last_nap = current_time
         sync()
-        cuda.memcpy_dtoh(self.result, self._result_b)
+        cuda.memcpy_dtoh(self.result, self.result_b)
         #print("result", self.result)
 
@@ -523 +496 @@
         Release resources associated with the kernel.
         """
-        self.q_input.release()
-        if self._result_b is not None:
-            self._result_b.free()
-            self._result_b = None
+        for p in self._need_release:
+            p.free()
+        self._need_release = []
 
     def __del__(self):
@@ -540 +512 @@
     Note: Maybe context.synchronize() is sufficient.
     """
-    # Create an event with which to synchronize.
+    #return # The following works in C++; don't know what pycuda is doing
+    # Create an event with which to synchronize
     done = cuda.Event()
 
@@ -546 +519 @@
     done.record()
 
-    # Make sure we don't hog resource while waiting to sync.
+    #line added to not hog resources
     while not done.query():
         time.sleep(0.01)
@@ -552 +525 @@
     # Block until the GPU executes the kernel.
     done.synchronize()
-
     # Clean up the event; I don't think they can be reused.
     del done

sasmodels/kerneldll.py

@@ -100 +100 @@
 # pylint: enable=unused-import
 
-# Compiler output is a byte stream that needs to be decode in python 3.
+# Compiler output is a byte stream that needs to be decode in python 3
 decode = (lambda s: s) if sys.version_info[0] < 3 else (lambda s: s.decode('utf8'))
 
@@ -115 +115 @@
         COMPILER = "tinycc"
     elif "VCINSTALLDIR" in os.environ:
-        # If vcvarsall.bat has been called, then VCINSTALLDIR is in the
-        # environment and we can use the MSVC compiler.  Otherwise, if
-        # tinycc is available then use it.  Otherwise, hope that mingw
-        # is available.
+        # If vcvarsall.bat has been called, then VCINSTALLDIR is in the environment
+        # and we can use the MSVC compiler.  Otherwise, if tinycc is available
+        # the use it.  Otherwise, hope that mingw is available.
         COMPILER = "msvc"
     else:
@@ -125 +124 @@
     COMPILER = "unix"
 
-ARCH = "" if ct.sizeof(ct.c_void_p) > 4 else "x86"  # 4 byte pointers on x86.
+ARCH = "" if ct.sizeof(ct.c_void_p) > 4 else "x86"  # 4 byte pointers on x86
 if COMPILER == "unix":
-    # Generic unix compile.
-    # On Mac users will need the X code command line tools installed.
+    # Generic unix compile
+    # On mac users will need the X code command line tools installed
     #COMPILE = "gcc-mp-4.7 -shared -fPIC -std=c99 -fopenmp -O2 -Wall %s -o %s -lm -lgomp"
     CC = "cc -shared -fPIC -std=c99 -O2 -Wall".split()
-    # Add OpenMP support if not running on a Mac.
+    # add openmp support if not running on a mac
     if sys.platform != "darwin":
-        # OpenMP seems to be broken on gcc 5.4.0 (ubuntu 16.04.9).
+        # OpenMP seems to be broken on gcc 5.4.0 (ubuntu 16.04.9)
         # Shut it off for all unix until we can investigate.
         #CC.append("-fopenmp")
@@ -145 +144 @@
     # vcomp90.dll on the path.  One may be found here:
     #       C:/Windows/winsxs/x86_microsoft.vc90.openmp*/vcomp90.dll
-    # Copy this to the python directory and uncomment the OpenMP COMPILE.
-    # TODO: Remove intermediate OBJ file created in the directory.
-    # TODO: Maybe don't use randomized name for the c file.
-    # TODO: Maybe ask distutils to find MSVC.
+    # Copy this to the python directory and uncomment the OpenMP COMPILE
+    # TODO: remove intermediate OBJ file created in the directory
+    # TODO: maybe don't use randomized name for the c file
+    # TODO: maybe ask distutils to find MSVC
     CC = "cl /nologo /Ox /MD /W3 /GS- /DNDEBUG".split()
     if "SAS_OPENMP" in os.environ:
@@ -173 +172 @@
 ALLOW_SINGLE_PRECISION_DLLS = True
 
-
 def compile(source, output):
     # type: (str, str) -> None
@@ -185 +183 @@
     logging.info(command_str)
     try:
-        # Need shell=True on windows to keep console box from popping up.
+        # need shell=True on windows to keep console box from popping up
         shell = (os.name == 'nt')
         subprocess.check_output(command, shell=shell, stderr=subprocess.STDOUT)
@@ -194 +192 @@
         raise RuntimeError("compile failed.  File is in %r"%source)
 
-
 def dll_name(model_info, dtype):
     # type: (ModelInfo, np.dtype) ->  str
@@ -205 +202 @@
     basename += ARCH + ".so"
 
-    # Hack to find precompiled dlls.
+    # Hack to find precompiled dlls
     path = joinpath(generate.DATA_PATH, '..', 'compiled_models', basename)
     if os.path.exists(path):
@@ -245 +242 @@
         raise ValueError("16 bit floats not supported")
     if dtype == F32 and not ALLOW_SINGLE_PRECISION_DLLS:
-        dtype = F64  # Force 64-bit dll.
-    # Note: dtype may be F128 for long double precision.
+        dtype = F64  # Force 64-bit dll
+    # Note: dtype may be F128 for long double precision
 
     dll = dll_path(model_info, dtype)
@@ -257 +254 @@
         need_recompile = dll_time < newest_source
     if need_recompile:
-        # Make sure the DLL path exists.
+        # Make sure the DLL path exists
         if not os.path.exists(SAS_DLL_PATH):
             os.makedirs(SAS_DLL_PATH)
@@ -266 +263 @@
             file_handle.write(source)
         compile(source=filename, output=dll)
-        # Comment the following to keep the generated C file.
-        # Note: If there is a syntax error then compile raises an error
+        # comment the following to keep the generated c file
+        # Note: if there is a syntax error then compile raises an error
         # and the source file will not be deleted.
         os.unlink(filename)
@@ -306 +303 @@
         self.dllpath = dllpath
         self._dll = None  # type: ct.CDLL
-        self._kernels = None  # type: List[Callable, Callable]
+        self._kernels = None # type: List[Callable, Callable]
         self.dtype = np.dtype(dtype)
 
@@ -341 +338 @@
         # type: (List[np.ndarray]) -> DllKernel
         q_input = PyInput(q_vectors, self.dtype)
-        # Note: DLL is lazy loaded.
+        # Note: pickle not supported for DllKernel
         if self._dll is None:
             self._load_dll()
@@ -361 +358 @@
         self._dll = None
 
-
 class DllKernel(Kernel):
     """
@@ -383 +379 @@
     def __init__(self, kernel, model_info, q_input):
         # type: (Callable[[], np.ndarray], ModelInfo, PyInput) -> None
-        dtype = q_input.dtype
+        #,model_info,q_input)
+        self.kernel = kernel
+        self.info = model_info
         self.q_input = q_input
-        self.kernel = kernel
-
-        # Attributes accessed from the outside.
+        self.dtype = q_input.dtype
         self.dim = '2d' if q_input.is_2d else '1d'
-        self.info = model_info
-        self.dtype = dtype
-
-        # Converter to translate input to target type.
-        self._as_dtype = (np.float32 if dtype == generate.F32
-                          else np.float64 if dtype == generate.F64
-                          else np.float128)
-
-        # Holding place for the returned value.
+        # leave room for f1/f2 results in case we need to compute beta for 1d models
         nout = 2 if self.info.have_Fq else 1
-        extra_q = 4  # Total weight, form volume, shell volume and R_eff.
-        self.result = np.empty(self.q_input.nq*nout + extra_q, dtype)
-
-    def _call_kernel(self, call_details, values, cutoff, magnetic,
-                     effective_radius_type):
-        # type: (CallDetails, np.ndarray, float, bool, int) -> np.ndarray
-
-        # Setup kernel function and arguments.
+        # +4 for total weight, shell volume, effective radius, form volume
+        self.result = np.empty(q_input.nq*nout + 4, self.dtype)
+        self.real = (np.float32 if self.q_input.dtype == generate.F32
+                     else np.float64 if self.q_input.dtype == generate.F64
+                     else np.float128)
+
+    def _call_kernel(self, call_details, values, cutoff, magnetic, effective_radius_type):
+        # type: (CallDetails, np.ndarray, np.ndarray, float, bool, int) -> np.ndarray
         kernel = self.kernel[1 if magnetic else 0]
-        kernel_args = [
-            self.q_input.nq,  # Number of inputs.
-            None,  # Placeholder for pd_start.
-            None,  # Placeholder for pd_stop.
-            call_details.buffer.ctypes.data,  # Problem definition.
-            values.ctypes.data,  # Parameter values.
-            self.q_input.q.ctypes.data,  # Q values.
-            self.result.ctypes.data,   # Result storage.
-            self._as_dtype(cutoff),  # Probability cutoff.
-            effective_radius_type,  # R_eff mode.
+        args = [
+            self.q_input.nq, # nq
+            None, # pd_start
+            None, # pd_stop pd_stride[MAX_PD]
+            call_details.buffer.ctypes.data, # problem
+            values.ctypes.data,  # pars
+            self.q_input.q.ctypes.data, # q
+            self.result.ctypes.data,   # results
+            self.real(cutoff), # cutoff
+            effective_radius_type, # cutoff
         ]
-
-        # Call kernel and retrieve results.
         #print("Calling DLL")
         #call_details.show(values)
         step = 100
-        # TODO: Do we need the explicit sleep like the OpenCL and CUDA loops?
         for start in range(0, call_details.num_eval, step):
             stop = min(start + step, call_details.num_eval)
-            kernel_args[1:3] = [start, stop]
-            kernel(*kernel_args) # type: ignore
+            args[1:3] = [start, stop]
+            kernel(*args) # type: ignore
 
     def release(self):
         # type: () -> None
         """
-        Release resources associated with the kernel.
+        Release any resources associated with the kernel.
         """
-        # TODO: OpenCL/CUDA allocate q_input in __init__ and free it in release.
-        # Should we be doing the same for DLL?
-        #self.q_input.release()
-        pass
-
-    def __del__(self):
-        # type: () -> None
-        self.release()
+        self.q_input.release()

sasmodels/kernelpy.py

@@ -33 +33 @@
 logger = logging.getLogger(__name__)
 
-
 class PyModel(KernelModel):
     """
@@ -39 +38 @@
     """
     def __init__(self, model_info):
-        # Make sure Iq is available and vectorized.
+        # Make sure Iq is available and vectorized
         _create_default_functions(model_info)
         self.info = model_info
@@ -54 +53 @@
         """
         pass
-
 
 class PyInput(object):
@@ -93 +91 @@
         self.q = None
 
-
 class PyKernel(Kernel):
     """
@@ -134 +131 @@
         parameter_vector = np.empty(len(partable.call_parameters)-2, 'd')
 
-        # Create views into the array to hold the arguments.
+        # Create views into the array to hold the arguments
         offset = 0
         kernel_args, volume_args = [], []
@@ -177 +174 @@
                         else (lambda mode: 1.0))
 
+
+
     def _call_kernel(self, call_details, values, cutoff, magnetic, effective_radius_type):
         # type: (CallDetails, np.ndarray, np.ndarray, float, bool) -> np.ndarray
@@ -196 +195 @@
         self.q_input.release()
         self.q_input = None
-
 
 def _loops(parameters,    # type: np.ndarray
@@ -256 +254 @@
         total = np.zeros(nq, 'd')
         for loop_index in range(call_details.num_eval):
-            # Update polydispersity parameter values.
+            # update polydispersity parameter values
             if p0_index == p0_length:
                 pd_index = (loop_index//pd_stride)%pd_length
@@ -267 +265 @@
             p0_index += 1
             if weight > cutoff:
-                # Call the scattering function.
+                # Call the scattering function
                 # Assume that NaNs are only generated if the parameters are bad;
                 # exclude all q for that NaN.  Even better would be to have an
@@ -275 +273 @@
                     continue
 
-                # Update value and norm.
+                # update value and norm
                 total += weight * Iq
                 weight_norm += weight
@@ -295 +293 @@
     any functions that are not already marked as vectorized.
     """
-    # Note: Must call create_vector_Iq before create_vector_Iqxy.
+    # Note: must call create_vector_Iq before create_vector_Iqxy
     _create_vector_Iq(model_info)
     _create_vector_Iqxy(model_info)

sasmodels/model_test.py

@@ -167 +167 @@
         # test using cuda if desired and available
         if 'cuda' in loaders and use_cuda():
-            test_name = "%s-cuda" % model_info.id
+            test_name = "%s-cuda"%model_name
             test_method_name = "test_%s_cuda" % model_info.id
             # Using dtype=None so that the models that are only

sasmodels/models/rpa.c

@@ -25 +25 @@
   double S0ba,Pbb,S0bb,Pbc,S0bc,Pbd,S0bd;
   double S0ca,S0cb,Pcc,S0cc,Pcd,S0cd;
-  //double S0da,S0db,S0dc;
+  double S0da,S0db,S0dc;
   double Pdd,S0dd;
   double Kaa,Kbb,Kcc;
   double Kba,Kca,Kcb;
-  //double Kda,Kdb,Kdc,Kdd;
+  double Kda,Kdb,Kdc,Kdd;
   double Zaa,Zab,Zac,Zba,Zbb,Zbc,Zca,Zcb,Zcc;
   double DenT,T11,T12,T13,T21,T22,T23,T31,T32,T33;
@@ -36 +36 @@
   double N11,N12,N13,N21,N22,N23,N31,N32,N33;
   double M11,M12,M13,M21,M22,M23,M31,M32,M33;
-  double S11,S12,S22,S23,S13,S33;
-  //double S21,S31,S32,S44; 
-  //double S14,S24,S34,S41,S42,S43;
+  double S11,S12,S13,S14,S21,S22,S23,S24;
+  double S31,S32,S33,S34,S41,S42,S43,S44;
   double Lad,Lbd,Lcd,Nav,Intg;
 
@@ -116 +115 @@
   S0cd=(Phicd*vcd*Ncd)*Pcd;
 
-  //S0da=S0ad;
-  //S0db=S0bd;
-  //S0dc=S0cd;
+  S0da=S0ad;
+  S0db=S0bd;
+  S0dc=S0cd;
   Pdd=2.0*(exp(-Xd)-1.0+Xd)/(Xd*Xd); // free D chain
   S0dd=N[3]*Phi[3]*v[3]*Pdd;
@@ -199 +198 @@
   S0ca=S0ac;
   S0cb=S0bc;
-  //S0da=S0ad;
-  //S0db=S0bd;
-  //S0dc=S0cd;
+  S0da=S0ad;
+  S0db=S0bd;
+  S0dc=S0cd;
 
   // self chi parameter is 0 ... of course
@@ -207 +206 @@
   Kbb=0.0;
   Kcc=0.0;
-  //Kdd=0.0;
+  Kdd=0.0;
 
   Kba=Kab;
   Kca=Kac;
   Kcb=Kbc;
-  //Kda=Kad;
-  //Kdb=Kbd;
-  //Kdc=Kcd;
+  Kda=Kad;
+  Kdb=Kbd;
+  Kdc=Kcd;
 
   Zaa=Kaa-Kad-Kad;
@@ -295 +294 @@
   S12= Q12*S0aa + Q22*S0ab + Q32*S0ac;
   S13= Q13*S0aa + Q23*S0ab + Q33*S0ac;
+  S14=-S11-S12-S13;
+  S21= Q11*S0ba + Q21*S0bb + Q31*S0bc;
   S22= Q12*S0ba + Q22*S0bb + Q32*S0bc;
   S23= Q13*S0ba + Q23*S0bb + Q33*S0bc;
+  S24=-S21-S22-S23;
+  S31= Q11*S0ca + Q21*S0cb + Q31*S0cc;
+  S32= Q12*S0ca + Q22*S0cb + Q32*S0cc;
   S33= Q13*S0ca + Q23*S0cb + Q33*S0cc;
-  //S21= Q11*S0ba + Q21*S0bb + Q31*S0bc;
-  //S31= Q11*S0ca + Q21*S0cb + Q31*S0cc;
-  //S32= Q12*S0ca + Q22*S0cb + Q32*S0cc;
-  //S44=S11+S22+S33+2.0*S12+2.0*S13+2.0*S23;
-  //S14=-S11-S12-S13;
-  //S24=-S21-S22-S23;
-  //S34=-S31-S32-S33;
-  //S41=S14;
-  //S42=S24;
-  //S43=S34;
+  S34=-S31-S32-S33;
+  S41=S14;
+  S42=S24;
+  S43=S34;
+  S44=S11+S22+S33+2.0*S12+2.0*S13+2.0*S23;
 
   //calculate contrast where L[i] is the scattering length of i and D is the matrix