Merge branch 'run-opencl-without-pycuda' into 'master'

Run opencl without pycuda

Closes #15

See merge request pycodegen/pystencils!77

pystencils/backends/opencl_backend.py

@@ -81,7 +81,10 @@ class OpenClSympyPrinter(CudaSympyPrinter):
     # For math functions, OpenCL is more similar to the C++ printer CustomSympyPrinter
     # since built-in math functions are generic.
     # In CUDA, you have to differentiate between `sin` and `sinf`
-    _print_math_func = CustomSympyPrinter._print_math_func
+    try:
+        _print_math_func = CustomSympyPrinter._print_math_func
+    except AttributeError:
+        pass
     _print_Pow = CustomSympyPrinter._print_Pow
 
     def _print_Function(self, expr):

pystencils/opencl/opencljit.py

@@ -41,8 +41,12 @@ def make_python_function(kernel_function_node, opencl_queue, opencl_ctx, argumen
     code += str(generate_c(kernel_function_node, dialect='opencl', custom_backend=custom_backend))
     options = []
     if USE_FAST_MATH:
-        options.append("-cl-unsafe-math-optimizations -cl-mad-enable -cl-fast-relaxed-math -cl-finite-math-only")
-    options.append("-I \"" + get_pystencils_include_path() + "\"")
+        options.append("-cl-unsafe-math-optimizations")
+        options.append("-cl-mad-enable")
+        options.append("-cl-fast-relaxed-math")
+        options.append("-cl-finite-math-only")
+    options.append("-I")
+    options.append(get_pystencils_include_path())
     mod = cl.Program(opencl_ctx, code).build(options=options)
     func = getattr(mod, kernel_function_node.function_name)
 

pystencils/transformations.py

@@ -1315,13 +1315,16 @@ def implement_interpolations(ast_node: ast.Node,
         substitutions = {i: to_texture_map[i.symbol.interpolator].at(
             [o for o in i.offsets]) for i in interpolation_accesses}
 
-        import pycuda.driver as cuda
-        for texture in substitutions.values():
-            if can_use_hw_interpolation(texture):
-                texture.filter_mode = cuda.filter_mode.LINEAR
-            else:
-                texture.filter_mode = cuda.filter_mode.POINT
-                texture.read_as_integer = True
+        try:
+            import pycuda.driver as cuda
+            for texture in substitutions.values():
+                if can_use_hw_interpolation(texture):
+                    texture.filter_mode = cuda.filter_mode.LINEAR
+                else:
+                    texture.filter_mode = cuda.filter_mode.POINT
+                    texture.read_as_integer = True
+        except Exception:
+            pass
 
         if isinstance(ast_node, AssignmentCollection):
             ast_node = ast_node.subs(substitutions)

pystencils_tests/test_opencl.py

 import numpy as np
 import pytest
-import sympy as sp
 
 import pystencils
+import sympy as sp
 from pystencils.backends.cuda_backend import CudaBackend
 from pystencils.backends.opencl_backend import OpenClBackend
 from pystencils.opencl.opencljit import make_python_function
@@ -40,6 +40,8 @@ def test_print_opencl():
 
 @pytest.mark.skipif(not HAS_OPENCL, reason="Test requires pyopencl")
 def test_opencl_jit_fixed_size():
+    pytest.importorskip('pycuda')
+
     z, y, x = pystencils.fields("z, y, x: [20,30]")
 
     assignments = pystencils.AssignmentCollection({
@@ -92,6 +94,8 @@ def test_opencl_jit_fixed_size():
 
 @pytest.mark.skipif(not HAS_OPENCL, reason="Test requires pyopencl")
 def test_opencl_jit():
+    pytest.importorskip('pycuda')
+
     z, y, x = pystencils.fields("z, y, x: [2d]")
 
     assignments = pystencils.AssignmentCollection({
@@ -144,6 +148,8 @@ def test_opencl_jit():
 
 @pytest.mark.skipif(not HAS_OPENCL, reason="Test requires pyopencl")
 def test_opencl_jit_with_parameter():
+    pytest.importorskip('pycuda')
+
     z, y, x = pystencils.fields("z, y, x: [2d]")
 
     a = sp.Symbol('a')
@@ -195,5 +201,35 @@ def test_opencl_jit_with_parameter():
     assert np.allclose(result_cuda, result_opencl)
 
 
-if __name__ == '__main__':
-    test_opencl_jit()
+@pytest.mark.skipif(not HAS_OPENCL, reason="Test requires pyopencl")
+def test_without_cuda():
+    z, y, x = pystencils.fields("z, y, x: [20,30]")
+
+    assignments = pystencils.AssignmentCollection({
+        z[0, 0]: x[0, 0] * sp.log(x[0, 0] * y[0, 0])
+    })
+
+    print(assignments)
+
+    ast = pystencils.create_kernel(assignments, target='gpu')
+
+    print(ast)
+
+    opencl_code = pystencils.show_code(ast, custom_backend=OpenClBackend())
+    print(opencl_code)
+
+    x_cpu = np.random.rand(20, 30)
+    y_cpu = np.random.rand(20, 30)
+    z_cpu = np.random.rand(20, 30)
+
+    import pyopencl.array as array
+    ctx = cl.create_some_context(0)
+    queue = cl.CommandQueue(ctx)
+
+    x = array.to_device(queue, x_cpu)
+    y = array.to_device(queue, y_cpu)
+    z = array.to_device(queue, z_cpu)
+
+    opencl_kernel = make_python_function(ast, queue, ctx)
+    assert opencl_kernel is not None
+    opencl_kernel(x=x, y=y, z=z)