void
BackendLLVM::llvm_assign_initial_value (const Symbol& sym)
{
// Don't write over connections! Connection values are written into
// our layer when the earlier layer is run, as part of its code. So
// we just don't need to initialize it here at all.
if (sym.valuesource() == Symbol::ConnectedVal &&
!sym.typespec().is_closure_based())
return;
if (sym.typespec().is_closure_based() && sym.symtype() == SymTypeGlobal)
return;
int arraylen = std::max (1, sym.typespec().arraylength());
// Closures need to get their storage before anything can be
// assigned to them. Unless they are params, in which case we took
// care of it in the group entry point.
if (sym.typespec().is_closure_based() &&
sym.symtype() != SymTypeParam && sym.symtype() != SymTypeOutputParam) {
llvm_assign_zero (sym);
return;
}
if ((sym.symtype() == SymTypeLocal || sym.symtype() == SymTypeTemp)
&& shadingsys().debug_uninit()) {
// Handle the "debug uninitialized values" case
bool isarray = sym.typespec().is_array();
int alen = isarray ? sym.typespec().arraylength() : 1;
llvm::Value *u = NULL;
if (sym.typespec().is_closure_based()) {
// skip closures
}
else if (sym.typespec().is_floatbased())
u = ll.constant (std::numeric_limits<float>::quiet_NaN());
else if (sym.typespec().is_int_based())
u = ll.constant (std::numeric_limits<int>::min());
else if (sym.typespec().is_string_based())
u = ll.constant (Strings::uninitialized_string);
if (u) {
for (int a = 0; a < alen; ++a) {
llvm::Value *aval = isarray ? ll.constant(a) : NULL;
for (int c = 0; c < (int)sym.typespec().aggregate(); ++c)
llvm_store_value (u, sym, 0, aval, c);
}
}
return;
}
if ((sym.symtype() == SymTypeLocal || sym.symtype() == SymTypeTemp) &&
sym.typespec().is_string_based()) {
// Strings are pointers. Can't take any chance on leaving
// local/tmp syms uninitialized.
llvm_assign_zero (sym);
return; // we're done, the parts below are just for params
}
ASSERT_MSG (sym.symtype() == SymTypeParam || sym.symtype() == SymTypeOutputParam,
"symtype was %d, data type was %s", (int)sym.symtype(), sym.typespec().c_str());
if (sym.has_init_ops() && sym.valuesource() == Symbol::DefaultVal) {
// Handle init ops.
build_llvm_code (sym.initbegin(), sym.initend());
} else if (! sym.lockgeom() && ! sym.typespec().is_closure()) {
// geometrically-varying param; memcpy its default value
TypeDesc t = sym.typespec().simpletype();
ll.op_memcpy (llvm_void_ptr (sym), ll.constant_ptr (sym.data()),
t.size(), t.basesize() /*align*/);
if (sym.has_derivs())
llvm_zero_derivs (sym);
} else {
// Use default value
int num_components = sym.typespec().simpletype().aggregate;
TypeSpec elemtype = sym.typespec().elementtype();
for (int a = 0, c = 0; a < arraylen; ++a) {
llvm::Value *arrind = sym.typespec().is_array() ? ll.constant(a) : NULL;
if (sym.typespec().is_closure_based())
continue;
for (int i = 0; i < num_components; ++i, ++c) {
// Fill in the constant val
llvm::Value* init_val = 0;
if (elemtype.is_floatbased())
init_val = ll.constant (((float*)sym.data())[c]);
else if (elemtype.is_string())
init_val = ll.constant (((ustring*)sym.data())[c]);
else if (elemtype.is_int())
init_val = ll.constant (((int*)sym.data())[c]);
ASSERT (init_val);
llvm_store_value (init_val, sym, 0, arrind, i);
}
}
if (sym.has_derivs())
llvm_zero_derivs (sym);
}
// Handle interpolated params.
// FIXME -- really, we shouldn't assign defaults or run init ops if
// the values are interpolated. The perf hit is probably small, since
// there are so few interpolated params, but we should come back and
// fix this later.
if ((sym.symtype() == SymTypeParam || sym.symtype() == SymTypeOutputParam)
&& ! sym.lockgeom()) {
std::vector<llvm::Value*> args;
args.push_back (sg_void_ptr());
args.push_back (ll.constant (sym.name()));
args.push_back (ll.constant (sym.typespec().simpletype()));
args.push_back (ll.constant ((int) sym.has_derivs()));
args.push_back (llvm_void_ptr (sym));
ll.call_function ("osl_bind_interpolated_param",
&args[0], args.size());
}
}
llvm::Type *
BackendLLVM::llvm_type_groupdata ()
{
// If already computed, return it
if (m_llvm_type_groupdata)
return m_llvm_type_groupdata;
std::vector<llvm::Type*> fields;
int offset = 0;
int order = 0;
if (llvm_debug() >= 2)
std::cout << "Group param struct:\n";
// First, add the array that tells if each layer has run. But only make
// slots for the layers that may be called/used.
if (llvm_debug() >= 2)
std::cout << " layers run flags: " << m_num_used_layers
<< " at offset " << offset << "\n";
int sz = (m_num_used_layers + 3) & (~3); // Round up to 32 bit boundary
fields.push_back (ll.type_array (ll.type_bool(), sz));
offset += sz * sizeof(bool);
++order;
// Now add the array that tells which userdata have been initialized,
// and the space for the userdata values.
int nuserdata = (int) group().m_userdata_names.size();
if (nuserdata) {
if (llvm_debug() >= 2)
std::cout << " userdata initialized flags: " << nuserdata
<< " at offset " << offset << ", field " << order << "\n";
ustring *names = & group().m_userdata_names[0];
TypeDesc *types = & group().m_userdata_types[0];
int *offsets = & group().m_userdata_offsets[0];
int sz = (nuserdata + 3) & (~3);
fields.push_back (ll.type_array (ll.type_bool(), sz));
offset += nuserdata * sizeof(bool);
++order;
for (int i = 0; i < nuserdata; ++i) {
TypeDesc type = types[i];
int n = type.numelements() * 3; // always make deriv room
type.arraylen = n;
fields.push_back (llvm_type (type));
// Alignment
int align = type.basesize();
offset = OIIO::round_to_multiple_of_pow2 (offset, align);
if (llvm_debug() >= 2)
std::cout << " userdata " << names[i] << ' ' << type
<< ", field " << order << ", offset " << offset << "\n";
offsets[i] = offset;
offset += int(type.size());
++order;
}
}
// For each layer in the group, add entries for all params that are
// connected or interpolated, and output params. Also mark those
// symbols with their offset within the group struct.
m_param_order_map.clear ();
for (int layer = 0; layer < group().nlayers(); ++layer) {
ShaderInstance *inst = group()[layer];
if (inst->unused())
continue;
FOREACH_PARAM (Symbol &sym, inst) {
TypeSpec ts = sym.typespec();
if (ts.is_structure()) // skip the struct symbol itself
continue;
const int arraylen = std::max (1, sym.typespec().arraylength());
const int derivSize = (sym.has_derivs() ? 3 : 1);
ts.make_array (arraylen * derivSize);
fields.push_back (llvm_type (ts));
// Alignment
size_t align = sym.typespec().is_closure_based() ? sizeof(void*) :
sym.typespec().simpletype().basesize();
if (offset & (align-1))
offset += align - (offset & (align-1));
if (llvm_debug() >= 2)
std::cout << " " << inst->layername()
<< " (" << inst->id() << ") " << sym.mangled()
<< " " << ts.c_str() << ", field " << order
<< ", size " << derivSize * int(sym.size())
<< ", offset " << offset << std::endl;
sym.dataoffset ((int)offset);
offset += derivSize* int(sym.size());
m_param_order_map[&sym] = order;
++order;
}
}
bool ImageInputWrap::read_scanline(int y, int z, TypeDesc format, object &buffer,
stride_t xstride=AutoStride)
{
void *write_buf = make_write_buffer(buffer, m_input->spec().width * m_input->spec().nchannels * format.basesize());
return m_input->read_scanline(y, z, format, write_buf, xstride);
}
// TESTME
bool ImageInputWrap::read_tile(int x, int y, int z, TypeDesc format,
object &buffer, stride_t xstride=AutoStride,
stride_t ystride=AutoStride,
stride_t zstride=AutoStride)
{
void *write_buf = make_write_buffer(buffer, m_input->spec().tile_pixels() * m_input->spec().nchannels * format.basesize());
return m_input->read_tile(x, y, z, format, write_buf,
xstride, ystride, zstride);
}
// The read_image method is a bit different from the c++ interface.
// "function" is a function which takes a float, and the
// PyProgressCallback function is called automatically.
bool ImageInputWrap::read_image(TypeDesc format, object &buffer,
stride_t xstride=AutoStride,
stride_t ystride=AutoStride,
stride_t zstride=AutoStride,
object function=object(handle<>(Py_None)))
{
void *write_buf = make_write_buffer(buffer, m_input->spec().image_pixels() * m_input->spec().nchannels * format.basesize());
if (function==handle<>(Py_None)) {
return m_input->read_image(format, write_buf, xstride, ystride, zstride,
NULL, NULL);
}
else {
return m_input->read_image(format, write_buf, xstride, ystride, zstride,
&PyProgressCallback, &function);
}
}