def _assign(self, value):
if isinstance(value, (int, float)):
# if we have a c or f contiguous array, then use the speedy driver kernel
if self.flags.forc and float(value) >= 0:
drv.memset_d16(self.gpudata,
Flexpt.flex_from_native(value, self.iwl),
self.size)
# otherwise use our copy kerel
else:
OpTreeNode.build("copy", value, None, out=self)
elif isinstance(value, FlexptArray):
if self.flags.forc and value.flags.forc and self.iwl == value.iwl:
drv.memcpy_dtod(self.gpudata, value.gpudata, self.nbytes)
else:
OpTreeNode.build("copy", value, None, out=self)
elif isinstance(value, OpTreeNode):
value.execute(out=self)
else:
raise TypeError("Invalid type for assignment: %s" % type(value))
return self
def ones(self, shape, iwl, allocator=drv.mem_alloc, order="C"):
"""
Returns an array of the given shape and dtype filled with 0's.
"""
result = FlexptArray(self, shape, iwl, allocator, order=order)
drv.memset_d16(result.gpudata, self.flex_from_native(1,iwl), result.size)
return result
def _assign(self, value):
if isinstance(value, (int, float)):
# if we have a contiguous array, then use the speedy driver kernel
if self.is_contiguous:
value = self.dtype.type(value)
if self.dtype.itemsize == 1:
drv.memset_d8( self.gpudata,
unpack_from('B', value)[0],
self.size)
elif self.dtype.itemsize == 2:
drv.memset_d16(self.gpudata,
unpack_from('H', value)[0],
self.size)
else:
drv.memset_d32(self.gpudata,
unpack_from('I', value)[0],
self.size)
# otherwise use our copy kerel
else:
OpTreeNode.build("assign", self, value)
elif isinstance(value, GPUTensor):
# TODO: add an is_binary_compat like function
if self.is_contiguous and value.is_contiguous and self.dtype == value.dtype:
drv.memcpy_dtod(self.gpudata, value.gpudata, self.nbytes)
else:
OpTreeNode.build("assign", self, value)
# collapse and execute an op tree as a kernel
elif isinstance(value, OpTreeNode):
OpTreeNode.build("assign", self, value)
# assign to numpy array (same as set())
elif isinstance(value, np.ndarray):
self.set(value)
else:
raise TypeError("Invalid type for assignment: %s" % type(value))
return self
def _assign(self, value):
if isinstance(value, (int, float)):
# if we have a contiguous array, then use the speedy driver kernel
if self.is_contiguous:
value = self.dtype.type(value)
if self.dtype.itemsize == 1:
drv.memset_d8(self.gpudata,
unpack_from('B', value)[0], self.size)
elif self.dtype.itemsize == 2:
drv.memset_d16(self.gpudata,
unpack_from('H', value)[0], self.size)
else:
drv.memset_d32(self.gpudata,
unpack_from('I', value)[0], self.size)
# otherwise use our copy kerel
else:
OpTreeNode.build("assign", self, value)
elif isinstance(value, GPUTensor):
# TODO: add an is_binary_compat like function
if self.is_contiguous and value.is_contiguous and self.dtype == value.dtype:
drv.memcpy_dtod(self.gpudata, value.gpudata, self.nbytes)
else:
OpTreeNode.build("assign", self, value)
# collapse and execute an op tree as a kernel
elif isinstance(value, OpTreeNode):
OpTreeNode.build("assign", self, value)
# assign to numpy array (same as set())
elif isinstance(value, np.ndarray):
self.set(value)
else:
raise TypeError("Invalid type for assignment: %s" % type(value))
return self
def _assign(self, value):
if isinstance(value, (int, float)):
# if we have a c or f contiguous array, then use the speedy driver kernel
if self.flags.forc and float(value) >= 0:
drv.memset_d16(self.gpudata, Flexpt.flex_from_native(value,self.iwl), self.size)
# otherwise use our copy kerel
else:
OpTreeNode.build("copy", value, None, out=self)
elif isinstance(value, FlexptArray):
if self.flags.forc and value.flags.forc and self.iwl == value.iwl:
drv.memcpy_dtod(self.gpudata, value.gpudata, self.nbytes)
else:
OpTreeNode.build("copy", value, None, out=self)
elif isinstance(value, OpTreeNode):
value.execute(out=self)
else:
raise TypeError("Invalid type for assignment: %s" % type(value))
return self
def _autotune_kernel(n, k, type_args, tune_args):
load_cnt = 0
array_ids = set()
kernel_args = list()
tune_args = list(tune_args)
# Setup some fake data to autotune on
# Perhaps tune on the real data but we'd need a custom memoize for that,
# And we wouldn't be able to use n,k abstracted sizing
data = drv.mem_alloc(n * k * 2)
drv.memset_d16(data, 0, n * k)
kernel_args.extend((data, k, 1, k))
for arg in type_args:
arg_type, arg_idx = arg[0:2]
# Array operands
if arg_type is flexpt_array.FlexptArray:
if arg_idx not in array_ids:
array_ids.add(arg_idx)
flags = tune_args.pop()
size = 1
if flags & 1 == 0:
size *= n
if flags & 2 == 0:
size *= k
data = drv.mem_alloc(size * 2)
drv.memset_d16(data, 0, size)
kernel_args.extend((data, k, 1, 15, flags))
load_cnt += 1
# Constant operands
elif arg_type is int:
kernel_args.extend((0, 15))
# Operations
elif arg[2]: # postop_convert_to_flex
kernel_args.append(15)
repeat = min(500, max(1, 8192**2 // (n * k)))
# warm up the gpu clocks so our timings are accurate
cur_ctx = drv.Context.get_current()
if cur_ctx not in _context_warmup_set:
_context_warmup_set.add(cur_ctx)
kernel = _get_compound_ew_kernel(0, type_args)
for r in range(repeat * 10):
kernel.prepared_call((n, 1, 1), (32, 1, 1), *kernel_args)
start = drv.Event()
end = drv.Event()
min_msecs = 99999999.9
min_blocks = 1
min_threads = 32
min_factor = 0
max_factor = 3 if load_cnt < 4 else 2
for unroll_factor in range(max_factor):
kernel = _get_compound_ew_kernel(unroll_factor, type_args)
unroll = 1 << unroll_factor
for threads in (32, 64, 128, 256):
for blocks in (1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024):
loads = blocks * threads * unroll
if loads > k and min_msecs != 99999999.9:
#print "skipping %d loads for %d" % (loads, k)
continue
loops = k // loads
if (loops > 8 or loops < 1) and min_msecs != 99999999.9:
print "skipping %d loops %d // %d" % (loops, k, loads)
continue
start.record()
for r in range(repeat):
kernel.prepared_call((n, blocks, 1), (threads, 1, 1),
*kernel_args)
end.record()
end.synchronize()
msecs = end.time_since(start)
if msecs < min_msecs:
min_msecs = msecs
min_factor = unroll_factor
min_blocks = blocks
min_threads = threads
#print "%4d %d %4d %3d %4d (%4d, %4d) %4d %.5f" % \
# (repeat, unroll, blocks, threads, loads, n, k, loops, msecs)
#print "Final: %d %4d %3d %.5f" % (1<<min_factor, min_blocks, min_threads, min_msecs)
return (min_factor, min_blocks, min_threads)
def _autotune_kernel(n, k, type_args, tune_args):
load_cnt = 0
array_ids = set()
kernel_args = list()
tune_args = list(tune_args)
# Setup some fake data to autotune on
# Perhaps tune on the real data but we'd need a custom memoize for that,
# And we wouldn't be able to use n,k abstracted sizing
data = drv.mem_alloc(n * k * 2)
drv.memset_d16(data, 0, n * k)
kernel_args.extend((data, k, 1, k))
for arg in type_args:
arg_type, arg_idx = arg[0:2]
# Array operands
if arg_type is flexpt_array.FlexptArray:
if arg_idx not in array_ids:
array_ids.add(arg_idx)
flags = tune_args.pop()
size = 1
if flags & 1 == 0:
size *= n
if flags & 2 == 0:
size *= k
data = drv.mem_alloc(size * 2)
drv.memset_d16(data, 0, size)
kernel_args.extend((data, k, 1, 15, flags))
load_cnt += 1
# Constant operands
elif arg_type is int:
kernel_args.extend((0,15))
# Operations
elif arg[2]: # postop_convert_to_flex
kernel_args.append(15)
repeat = min(500, max(1, 8192**2 // (n * k)))
# warm up the gpu clocks so our timings are accurate
cur_ctx = drv.Context.get_current()
if cur_ctx not in _context_warmup_set:
_context_warmup_set.add(cur_ctx)
kernel = _get_compound_ew_kernel(0, type_args)
for r in range(repeat * 10):
kernel.prepared_call((n,1,1), (32,1,1), *kernel_args)
start = drv.Event()
end = drv.Event()
min_msecs = 99999999.9
min_blocks = 1
min_threads = 32
min_factor = 0
max_factor = 3 if load_cnt < 4 else 2
for unroll_factor in range(max_factor):
kernel = _get_compound_ew_kernel(unroll_factor, type_args)
unroll = 1 << unroll_factor
for threads in (32,64,128,256):
for blocks in (1,2,4,8,16,32,64,128,256,512,1024):
loads = blocks * threads * unroll
if loads > k and min_msecs != 99999999.9:
#print "skipping %d loads for %d" % (loads, k)
continue
loops = k // loads
if (loops > 8 or loops < 1) and min_msecs != 99999999.9:
print "skipping %d loops %d // %d" % (loops, k, loads)
continue
start.record()
for r in range(repeat):
kernel.prepared_call((n,blocks,1), (threads,1,1), *kernel_args)
end.record()
end.synchronize()
msecs = end.time_since(start)
if msecs < min_msecs:
min_msecs = msecs
min_factor = unroll_factor
min_blocks = blocks
min_threads = threads
#print "%4d %d %4d %3d %4d (%4d, %4d) %4d %.5f" % \
# (repeat, unroll, blocks, threads, loads, n, k, loops, msecs)
#print "Final: %d %4d %3d %.5f" % (1<<min_factor, min_blocks, min_threads, min_msecs)
return (min_factor, min_blocks, min_threads)
