Changes in sasmodels/core.py [2dcd6e7:b0de252] in sasmodels

File:

• sasmodels/core.py (modified) (7 diffs)

sasmodels/core.py

@@ -21 +21 @@
 from . import mixture
 from . import kernelpy
+from . import kernelcuda
 from . import kernelcl
 from . import kerneldll
@@ -210 +211 @@
         #print("building dll", numpy_dtype)
         return kerneldll.load_dll(source['dll'], model_info, numpy_dtype)
+    elif platform == "cuda":
+        return kernelcuda.GpuModel(source, model_info, numpy_dtype, fast=fast)
     else:
         #print("building ocl", numpy_dtype)
@@ -245 +248 @@
     # type: (ModelInfo, str, str) -> (np.dtype, bool, str)
     """
-    Interpret dtype string, returning np.dtype and fast flag.
+    Interpret dtype string, returning np.dtype, fast flag and platform.
 
     Possible types include 'half', 'single', 'double' and 'quad'.  If the
@@ -253 +256 @@
     default for the model and platform.
 
-    Platform preference can be specfied ("ocl" vs "dll"), with the default
-    being OpenCL if it is availabe.  If the dtype name ends with '!' then
-    platform is forced to be DLL rather than OpenCL.
+    Platform preference can be specfied ("ocl", "cuda", "dll"), with the
+    default being OpenCL or CUDA if available, otherwise DLL.  If the dtype
+    name ends with '!' then platform is forced to be DLL rather than GPU.
+    The default platform is set by the environment variable SAS_OPENCL,
+    SAS_OPENCL=driver:device for OpenCL, SAS_OPENCL=cuda:device for CUDA
+    or SAS_OPENCL=none for DLL.
 
     This routine ignores the preferences within the model definition.  This
@@ -268 +274 @@
     if platform is None:
         platform = "ocl"
-    if not kernelcl.use_opencl() or not model_info.opencl:
-        platform = "dll"
 
     # Check if type indicates dll regardless of which platform is given
@@ -275 +279 @@
         platform = "dll"
         dtype = dtype[:-1]
+
+    # Make sure model allows opencl/gpu
+    if not model_info.opencl:
+        platform = "dll"
+
+    # Make sure opencl is available, or fallback to cuda then to dll
+    if platform == "ocl" and not kernelcl.use_opencl():
+        platform = "cuda" if kernelcuda.use_cuda() else "dll"
 
     # Convert special type names "half", "fast", and "quad"
@@ -285 +297 @@
         dtype = "float16"
 
-    # Convert dtype string to numpy dtype.
+    # Convert dtype string to numpy dtype.  Use single precision for GPU
+    # if model allows it, otherwise use double precision.
     if dtype is None or dtype == "default":
-        numpy_dtype = (generate.F32 if platform == "ocl" and model_info.single
+        numpy_dtype = (generate.F32 if model_info.single and platform in ("ocl", "cuda")
                        else generate.F64)
     else:
         numpy_dtype = np.dtype(dtype)
 
-    # Make sure that the type is supported by opencl, otherwise use dll
+    # Make sure that the type is supported by GPU, otherwise use dll
     if platform == "ocl":
         env = kernelcl.environment()
-        if not env.has_type(numpy_dtype):
-            platform = "dll"
-            if dtype is None:
-                numpy_dtype = generate.F64
+    elif platform == "cuda":
+        env = kernelcuda.environment()
+    else:
+        env = None
+    if env is not None and not env.has_type(numpy_dtype):
+        platform = "dll"
+        if dtype is None:
+            numpy_dtype = generate.F64
 
     return numpy_dtype, fast, platform