Python/OpenCL

PyOpenCL

Passing data to kernel functions.

Prefix

OpenCL	Python
__global const long* data	# Read Only py_param = [1, 2, 3, 4, 5, 6] cl_param = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=np.int64(py_param))
__global long* data	# Read Write py_param = np.int64([1, 2, 3, 4, 5, 6]) cl_param = cl.Buffer(ctx, mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=py_param) # Write Only py_param = np.int64([0, 0, 0, 0, 0]) cl_param = cl.Buffer(ctx, mf.WRITE_ONLY, py_param.nbytes)
__private long i	cl_param = np.int64(9999)

Type Mapping

OpenCL	Python
float	np.float32
double	np.float64
long	np.int64
int	np.int32
short	np.int16
char	np.int8 or bytearray

See Also:

• Scalar Data Types

Sample

import os
import numpy as np
import pyopencl as cl

# Kernel code.
CL_INC = '''
__kernel void inc_f32(__global const float *a_g, __global float *res_g)
{
  int gid = get_global_id(0);
  res_g[gid] = a_g[gid] + 1.0;
}

__kernel void inc_f64(__global const double *a_g, __global double *res_g)
{
  int gid = get_global_id(0);
  res_g[gid] = a_g[gid] + 1.0;
}

__kernel void inc_i32(__global const int *a_g, __global int *res_g)
{
  int gid = get_global_id(0);
  res_g[gid] = a_g[gid] + 1;
}

__kernel void inc_i64(__global const long *a_g, __global long *res_g)
{
  int gid = get_global_id(0);
  res_g[gid] = a_g[gid] + 1;
}
'''

# Build
os.environ['PYOPENCL_CTX'] = '0:1' # Apple > HD Graphics 4000 of Macbook Air 2012 mid
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
mf = cl.mem_flags
prg = cl.Program(ctx, CL_INC).build()

# Shorthand to call kernel function.
def call_kernel(kernel_name, *args):
    in_np = args[0]
    out_np = np.empty_like(in_np)
    in_cl = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=in_np)
    out_cl = cl.Buffer(ctx, mf.WRITE_ONLY, in_np.nbytes)
    getattr(prg, kernel_name)(queue, in_np.shape, None, in_cl, out_cl)
    cl.enqueue_copy(queue, out_np, out_cl)
    return out_np

# Test inc_f32()
in_np = np.array([1.0, 2.0, 3.0, 4.0, 5.0]).astype(np.float32)
out_np = call_kernel('inc_f32', in_np)
print('Results of inc_f32()')
print(in_np)
print(out_np)
print()

# Test inc_f64()
in_np = np.array([1.0, 2.0, 3.0, 4.0, 5.0])
out_np = call_kernel('inc_f64', in_np)
print('Results of inc_f64()')
print(in_np)
print(out_np)
print()

# Test inc_i32()
in_np = np.array([1, 2, 3, 4, 5]).astype(np.int32)
out_np = call_kernel('inc_i32', in_np)
print('Results of inc_i32()')
print(in_np)
print(out_np)
print()

# Test inc_i64()
in_np = np.array([1, 2, 3, 4, 5])
out_np = call_kernel('inc_i64', in_np)
print('Results of inc_i64()')
print(in_np)
print(out_np)
print()

numpy

• https://docs.scipy.org/doc/numpy/user/quickstart.html

Data Types

import numpy as np

# Default int type is int64
a = np.array([1, 2, 3])
print(type(a[0]).__name__)

# Default float type is float64
b = np.array([1.0, 2.0, 3.0])
print(type(b[0]).__name__)

# Force int32
c = np.array([1, 2, 3]).astype(np.int32)
print(type(c[0]).__name__)

# Force float32
d = np.array([1.0, 2.0, 3.0]).astype(np.float32)
print(type(d[0]).__name__)