def test_inline_rsqrt(self):
mod = Module.new(__name__)
fnty = Type.function(Type.void(), [Type.pointer(Type.float())])
fn = mod.add_function(fnty, "cu_rsqrt")
bldr = Builder.new(fn.append_basic_block("entry"))
rsqrt_approx_fnty = Type.function(Type.float(), [Type.float()])
inlineasm = InlineAsm.get(rsqrt_approx_fnty, "rsqrt.approx.f32 $0, $1;", "=f,f", side_effect=True)
val = bldr.load(fn.args[0])
res = bldr.call(inlineasm, [val])
bldr.store(res, fn.args[0])
bldr.ret_void()
# generate ptx
nvvm.fix_data_layout(mod)
nvvm.set_cuda_kernel(fn)
nvvmir = str(mod)
ptx = nvvm.llvm_to_ptx(nvvmir)
self.assertTrue("rsqrt.approx.f32" in str(ptx))
def test_inline_rsqrt(self):
mod = Module.new(__name__)
fnty = Type.function(Type.void(), [Type.pointer(Type.float())])
fn = mod.add_function(fnty, 'cu_rsqrt')
bldr = Builder.new(fn.append_basic_block('entry'))
rsqrt_approx_fnty = Type.function(Type.float(), [Type.float()])
inlineasm = InlineAsm.get(rsqrt_approx_fnty,
'rsqrt.approx.f32 $0, $1;',
'=f,f', side_effect=True)
val = bldr.load(fn.args[0])
res = bldr.call(inlineasm, [val])
bldr.store(res, fn.args[0])
bldr.ret_void()
# generate ptx
nvvm.fix_data_layout(mod)
nvvm.set_cuda_kernel(fn)
nvvmir = str(mod)
ptx = nvvm.llvm_to_ptx(nvvmir)
self.assertTrue('rsqrt.approx.f32' in str(ptx))
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()