def get_dev_no_sync_size(device):
from pyopencl.characterize import get_simd_group_size
result = get_simd_group_size(device, out_type_size)
if result is None:
from warnings import warn
warn("Reduction might be unnecessarily slow: "
"can't query SIMD group size")
return 1
return result
def get_dev_no_sync_size(device):
from pyopencl.characterize import get_simd_group_size
result = get_simd_group_size(device, out_type_size)
if result is None:
from warnings import warn
warn("Reduction might be unnecessarily slow: "
"can't query SIMD group size")
return 1
return result
def build_preamble_for_context(ctx,preamble=""):
warp_size=clchar.get_simd_group_size(ctx.devices[0],0)
if(warp_size >= 32):
preamble += """
#define WARPSPEED
#define NOWARPBLOCK
"""
else:
preamble += """
#define NOWARPBLOCK barrier(CLK_LOCAL_MEM_FENCE);
"""
return preamble
def build_preamble_for_context(ctx, preamble=""):
warp_size = clchar.get_simd_group_size(ctx.devices[0], 0)
if (warp_size >= 32 and False):
preamble += """
#define WARPSPEED
#define NOWARPBLOCK
"""
else:
preamble += """
#define NOWARPBLOCK barrier(CLK_LOCAL_MEM_FENCE);
"""
return preamble
