def insert_addrspace_conv(self, builder, ptr, addrspace):
"""
Perform addrspace conversion according to the NVVM spec
"""
lmod = builder.module
base_type = ptr.type.pointee
conv = nvvmutils.insert_addrspace_conv(lmod, base_type, addrspace)
return builder.call(conv, [ptr])
def make_constant_array(self, builder, aryty, arr):
"""
Unlike the parent version. This returns a a pointer in the constant
addrspace.
"""
lmod = builder.module
constvals = [
self.get_constant(types.byte, i)
for i in iter(arr.tobytes(order='A'))
]
constaryty = ir.ArrayType(ir.IntType(8), len(constvals))
constary = ir.Constant(constaryty, constvals)
addrspace = nvvm.ADDRSPACE_CONSTANT
gv = cgutils.add_global_variable(lmod,
constary.type,
"_cudapy_cmem",
addrspace=addrspace)
gv.linkage = 'internal'
gv.global_constant = True
gv.initializer = constary
# Preserve the underlying alignment
lldtype = self.get_data_type(aryty.dtype)
align = self.get_abi_sizeof(lldtype)
gv.align = 2**(align - 1).bit_length()
# Convert to generic address-space
conv = nvvmutils.insert_addrspace_conv(lmod, ir.IntType(8), addrspace)
addrspaceptr = gv.bitcast(ir.PointerType(ir.IntType(8), addrspace))
genptr = builder.call(conv, [addrspaceptr])
# Create array object
ary = self.make_array(aryty)(self, builder)
kshape = [self.get_constant(types.intp, s) for s in arr.shape]
kstrides = [self.get_constant(types.intp, s) for s in arr.strides]
self.populate_array(ary,
data=builder.bitcast(genptr, ary.data.type),
shape=kshape,
strides=kstrides,
itemsize=ary.itemsize,
parent=ary.parent,
meminfo=None)
return ary._getvalue()
def _generic_array(context, builder, shape, dtype, symbol_name, addrspace,
can_dynsized=False):
elemcount = reduce(operator.mul, shape)
lldtype = context.get_data_type(dtype)
laryty = Type.array(lldtype, elemcount)
if addrspace == nvvm.ADDRSPACE_LOCAL:
# Special case local addrespace allocation to use alloca
# NVVM is smart enough to only use local memory if no register is
# available
dataptr = cgutils.alloca_once(builder, laryty, name=symbol_name)
else:
lmod = builder.module
# Create global variable in the requested address-space
gvmem = lmod.add_global_variable(laryty, symbol_name, addrspace)
# Specify alignment to avoid misalignment bug
align = context.get_abi_sizeof(lldtype)
# Alignment is required to be a power of 2 for shared memory. If it is
# not a power of 2 (e.g. for a Record array) then round up accordingly.
gvmem.align = 1 << (align - 1 ).bit_length()
if elemcount <= 0:
if can_dynsized: # dynamic shared memory
gvmem.linkage = lc.LINKAGE_EXTERNAL
else:
raise ValueError("array length <= 0")
else:
## Comment out the following line to workaround a NVVM bug
## which generates a invalid symbol name when the linkage
## is internal and in some situation.
## See _get_unique_smem_id()
# gvmem.linkage = lc.LINKAGE_INTERNAL
gvmem.initializer = lc.Constant.undef(laryty)
other_supported_type = isinstance(dtype, (types.Record, types.Boolean))
if dtype not in types.number_domain and not other_supported_type:
raise TypeError("unsupported type: %s" % dtype)
# Convert to generic address-space
conv = nvvmutils.insert_addrspace_conv(lmod, Type.int(8), addrspace)
addrspaceptr = gvmem.bitcast(Type.pointer(Type.int(8), addrspace))
dataptr = builder.call(conv, [addrspaceptr])
return _make_array(context, builder, dataptr, dtype, shape)
def ptx_cmem_arylike(context, builder, sig, args):
lmod = builder.module
[arr] = args
aryty = sig.return_type
constvals = [
context.get_constant(types.byte, i)
for i in iter(arr.tobytes(order="A"))
]
constary = lc.Constant.array(Type.int(8), constvals)
addrspace = nvvm.ADDRSPACE_CONSTANT
gv = lmod.add_global_variable(constary.type,
name="_cudapy_cmem",
addrspace=addrspace)
gv.linkage = lc.LINKAGE_INTERNAL
gv.global_constant = True
gv.initializer = constary
# Preserve the underlying alignment
lldtype = context.get_data_type(aryty.dtype)
align = context.get_abi_sizeof(lldtype)
gv.align = 2**(align - 1).bit_length()
# Convert to generic address-space
conv = nvvmutils.insert_addrspace_conv(lmod, Type.int(8), addrspace)
addrspaceptr = gv.bitcast(Type.pointer(Type.int(8), addrspace))
genptr = builder.call(conv, [addrspaceptr])
# Create array object
ary = context.make_array(aryty)(context, builder)
kshape = [context.get_constant(types.intp, s) for s in arr.shape]
kstrides = [context.get_constant(types.intp, s) for s in arr.strides]
context.populate_array(
ary,
data=builder.bitcast(genptr, ary.data.type),
shape=cgutils.pack_array(builder, kshape),
strides=cgutils.pack_array(builder, kstrides),
itemsize=ary.itemsize,
parent=ary.parent,
meminfo=None,
)
return ary._getvalue()
def _generic_array(context,
builder,
shape,
dtype,
symbol_name,
addrspace,
can_dynsized=False):
elemcount = reduce(operator.mul, shape, 1)
# Check for valid shape for this type of allocation.
# Only 1d arrays can be dynamic.
dynamic_smem = elemcount <= 0 and can_dynsized and len(shape) == 1
if elemcount <= 0 and not dynamic_smem:
raise ValueError("array length <= 0")
# Check that we support the requested dtype
other_supported_type = isinstance(dtype, (types.Record, types.Boolean))
if dtype not in types.number_domain and not other_supported_type:
raise TypeError("unsupported type: %s" % dtype)
lldtype = context.get_data_type(dtype)
laryty = Type.array(lldtype, elemcount)
if addrspace == nvvm.ADDRSPACE_LOCAL:
# Special case local address space allocation to use alloca
# NVVM is smart enough to only use local memory if no register is
# available
dataptr = cgutils.alloca_once(builder, laryty, name=symbol_name)
else:
lmod = builder.module
# Create global variable in the requested address space
gvmem = lmod.add_global_variable(laryty, symbol_name, addrspace)
# Specify alignment to avoid misalignment bug
align = context.get_abi_sizeof(lldtype)
# Alignment is required to be a power of 2 for shared memory. If it is
# not a power of 2 (e.g. for a Record array) then round up accordingly.
gvmem.align = 1 << (align - 1).bit_length()
if dynamic_smem:
gvmem.linkage = lc.LINKAGE_EXTERNAL
else:
## Comment out the following line to workaround a NVVM bug
## which generates a invalid symbol name when the linkage
## is internal and in some situation.
## See _get_unique_smem_id()
# gvmem.linkage = lc.LINKAGE_INTERNAL
gvmem.initializer = lc.Constant.undef(laryty)
# Convert to generic address-space
conv = nvvmutils.insert_addrspace_conv(lmod, Type.int(8), addrspace)
addrspaceptr = gvmem.bitcast(Type.pointer(Type.int(8), addrspace))
dataptr = builder.call(conv, [addrspaceptr])
targetdata = _get_target_data(context)
lldtype = context.get_data_type(dtype)
itemsize = lldtype.get_abi_size(targetdata)
# Compute strides
laststride = itemsize
rstrides = []
for i, lastsize in enumerate(reversed(shape)):
rstrides.append(laststride)
laststride *= lastsize
strides = [s for s in reversed(rstrides)]
kstrides = [context.get_constant(types.intp, s) for s in strides]
# Compute shape
if dynamic_smem:
# Compute the shape based on the dynamic shared memory configuration.
# Unfortunately NVVM does not provide an intrinsic for the
# %dynamic_smem_size register, so we must read it using inline
# assembly.
get_dynshared_size = InlineAsm.get(Type.function(Type.int(), []),
"mov.u32 $0, %dynamic_smem_size;",
'=r',
side_effect=True)
dynsmem_size = builder.zext(builder.call(get_dynshared_size, []),
Type.int(width=64))
# Only 1-D dynamic shared memory is supported so the following is a
# sufficient construction of the shape
kitemsize = context.get_constant(types.intp, itemsize)
kshape = [builder.udiv(dynsmem_size, kitemsize)]
else:
kshape = [context.get_constant(types.intp, s) for s in shape]
# Create array object
ndim = len(shape)
aryty = types.Array(dtype=dtype, ndim=ndim, layout='C')
ary = context.make_array(aryty)(context, builder)
context.populate_array(ary,
data=builder.bitcast(dataptr, ary.data.type),
shape=kshape,
strides=kstrides,
itemsize=context.get_constant(types.intp, itemsize),
meminfo=None)
return ary._getvalue()
