Changeset 5d4777d in sasmodels for sasmodels/gpu.py

Timestamp:

Sep 1, 2014 11:24:38 PM (10 years ago)

Author:

Paul Kienzle <pkienzle@…>

Branches:

master, core_shell_microgels, costrafo411, magnetic_model, release_v0.94, release_v0.95, ticket-1257-vesicle-product, ticket_1156, ticket_1265_superball, ticket_822_more_unit_tests

Children:

f4cf580

Parents:

ff7119b

Message:

reorganize, check and update models

File:

• sasmodels/gpu.py (modified) (9 diffs)

sasmodels/gpu.py

@@ -25 +25 @@
 
 """
-import warnings
-
 import numpy as np
 import pyopencl as cl
@@ -32 +30 @@
 
 from . import gen
-
-from .gen import F32, F64
 
 F32_DEFS = """\
@@ -52 +48 @@
 # larger than necessary given that cost grows as npts^k where k is the number
 # of polydisperse parameters.
-MAX_LOOPS = 1024
+MAX_LOOPS = 2048
+
+def load_model(kernel_module, dtype="single"):
+    """
+    Load the OpenCL model defined by *kernel_module*.
+
+    Access to the OpenCL device is delayed until the kernel is called
+    so models can be defined without using too many resources.
+    """
+    source, info = gen.make(kernel_module)
+    ## for debugging, save source to a .cl file, edit it, and reload as model
+    open(info['name']+'.cl','w').write(source)
+    #source = open(info['name']+'.cl','r').read()
+    return GpuModel(source, info, dtype)
 
 ENV = None
@@ -103 +112 @@
     """
     dtype = np.dtype(dtype)
-    if dtype==F64 and not all(has_double(d) for d in context.devices):
+    if dtype==gen.F64 and not all(has_double(d) for d in context.devices):
         raise RuntimeError("Double precision not supported for devices")
 
-    header = F64_DEFS if dtype == F64 else F32_DEFS
+    header = F64_DEFS if dtype == gen.F64 else F32_DEFS
     # Note: USE_SINCOS makes the intel cpu slower under opencl
     if context.devices[0].type == cl.device_type.GPU:
@@ -158 +167 @@
     is an optional extension which may not be available on all devices.
     """
-    def __init__(self, source, info, dtype=F32):
+    def __init__(self, source, info, dtype=gen.F32):
         self.info = info
         self.source = source
@@ -221 +230 @@
     buffer will be released when the data object is freed.
     """
-    def __init__(self, q_vectors, dtype=F32):
+    def __init__(self, q_vectors, dtype=gen.F32):
         env = environment()
         self.nq = q_vectors[0].size
@@ -273 +282 @@
         env = environment()
         self.loops_b = [cl.Buffer(env.context, mf.READ_WRITE,
-                                  MAX_LOOPS*input.dtype.itemsize)
+                                  2*MAX_LOOPS*input.dtype.itemsize)
                         for _ in env.queues]
         self.res_b = [cl.Buffer(env.context, mf.READ_WRITE,
@@ -281 +290 @@
 
     def __call__(self, pars, pd_pars, cutoff=1e-5):
-        real = np.float32 if self.input.dtype == F32 else np.float64
+        real = np.float32 if self.input.dtype == gen.F32 else np.float64
         fixed = [real(p) for p in pars]
         cutoff = real(cutoff)
         loops = np.hstack(pd_pars)
         loops = np.ascontiguousarray(loops.T, self.input.dtype).flatten()
-        loops_N = [np.uint32(len(p[0])) for p in pd_pars]
+        Nloops = [np.uint32(len(p[0])) for p in pd_pars]
+        #print "loops",Nloops, loops
 
         #import sys; print >>sys.stderr,"opencl eval",pars
         #print "opencl eval",pars
-        if len(loops) > MAX_LOOPS:
+        if len(loops) > 2*MAX_LOOPS:
             raise ValueError("too many polydispersity points")
         device_num = 0
@@ -300 +310 @@
         #ctx = environment().context
         #loops_bi = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=loops)
-        args = self.input.q_buffers + [res_bi,loops_bi,loops_l,cutoff] + fixed + loops_N
+        args = self.input.q_buffers + [res_bi,loops_bi,loops_l,cutoff] + fixed + Nloops
         self.kernel(queuei, self.input.global_size, None, *args)
         cl.enqueue_copy(queuei, self.res, res_bi)