AuroraMiddleware/SPIRV-Cross
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

MSL: Handle flattening of patch block outputs as well.

Browse Source 
Always propagate InterfaceMember decoration.
This commit is contained in:
Hans-Kristian Arntzen 2021-04-07 17:02:30 +02:00
parent 8e2dbe0d38
commit 425e968720
2 changed files with 45 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

77
spirv_msl.cpp
View File

@ -2561,7 +2561,7 @@ void CompilerMSL::add_composite_member_variable_to_interface_block(StorageClass
set_extended_member_decoration(ib_type.self, ib_mbr_idx, SPIRVCrossDecorationInterfaceMemberIndex, mbr_idx);
// Unflatten or flatten from [[stage_in]] or [[stage_out]] as appropriate.
if (!meta.strip_array)
if (!meta.strip_array && meta.allow_local_declaration)
{
switch (storage)
{
@ -2662,7 +2662,7 @@ void CompilerMSL::add_plain_member_variable_to_interface_block(StorageClass stor
// so redirect to qualified name.
set_member_qualified_name(var_type.self, mbr_idx, qual_var_name);
}
else if (!meta.strip_array)
else if (!meta.strip_array && meta.allow_local_declaration)
{
// Unflatten or flatten from [[stage_in]] or [[stage_out]] as appropriate.
switch (storage)
@ -2751,8 +2751,7 @@ void CompilerMSL::add_plain_member_variable_to_interface_block(StorageClass stor
{
entry_func.fixup_hooks_in.push_back([=, &var]() {
auto &type = this->get<SPIRType>(var.basetype);
uint32_t index = get_extended_decoration(var.self, SPIRVCrossDecorationInterfaceMemberIndex);
index += mbr_idx;
uint32_t index = get_extended_member_decoration(var.self, mbr_idx, SPIRVCrossDecorationInterfaceMemberIndex);
AccessChainMeta chain_meta;
auto constant_chain = access_chain_internal(var.initializer, &builtin_invocation_id_id, 1, 0, &chain_meta);
@ -2943,15 +2942,8 @@ void CompilerMSL::add_variable_to_interface_block(StorageClass storage, const st
if (var_type.basetype == SPIRType::Struct)
{
bool block_requires_split_members = !capture_output_to_buffer || is_block || storage == StorageClassInput;
bool needs_local_declaration =
!is_builtin_type(var_type) && block_requires_split_members && !meta.strip_array;
// Fixing up loads and stores in TESC is impossible since the memory is group shared either via
// device (not masked) or threadgroup (masked) storage classes.
if (get_execution_model() == ExecutionModelTessellationControl && storage == StorageClassOutput)
needs_local_declaration = false;
bool block_requires_flattening = is_block || !capture_output_to_buffer || storage == StorageClassInput;
bool needs_local_declaration = !is_builtin && block_requires_flattening && meta.allow_local_declaration;
if (needs_local_declaration)
{
@ -2964,7 +2956,7 @@ void CompilerMSL::add_variable_to_interface_block(StorageClass storage, const st
emit_local_masked_variable(var);
}
if (!block_requires_split_members)
if (!block_requires_flattening)
{
// In Metal tessellation shaders, the interface block itself is arrayed. This makes things
// very complicated, since stage-in structures in MSL don't support nested structures.
@ -3095,10 +3087,13 @@ void CompilerMSL::fix_up_interface_member_indices(StorageClass storage, uint32_t
auto &var = get<SPIRVariable>(var_id);
auto &type = get_variable_element_type(var);
if (storage == StorageClassInput && type.basetype == SPIRType::Struct)
{
uint32_t mbr_idx = get_extended_member_decoration(ib_type_id, i, SPIRVCrossDecorationInterfaceMemberIndex);
uint32_t mbr_idx = uint32_t(-1);
if (type.basetype == SPIRType::Struct)
mbr_idx = get_extended_member_decoration(ib_type_id, i, SPIRVCrossDecorationInterfaceMemberIndex);
if (mbr_idx != uint32_t(-1))
{
// Only set the lowest InterfaceMemberIndex for each variable member.
// IB struct members will be emitted in-order w.r.t. interface member index.
if (!has_extended_member_decoration(var_id, mbr_idx, SPIRVCrossDecorationInterfaceMemberIndex))
@ -3399,7 +3394,11 @@ uint32_t CompilerMSL::add_interface_block(StorageClass storage, bool patch)
(get_execution_model() == ExecutionModelTessellationEvaluation && storage == StorageClassInput)) &&
!patch;
// Fixing up flattened stores in TESC is impossible since the memory is group shared either via
// device (not masked) or threadgroup (masked) storage classes and it's race condition city.
meta.strip_array = strip_array;
meta.allow_local_declaration = !strip_array && !(get_execution_model() == ExecutionModelTessellationControl &&
storage == StorageClassOutput);
add_variable_to_interface_block(storage, ib_var_ref, ib_type, *p_var, meta);
}
@ -3452,6 +3451,7 @@ uint32_t CompilerMSL::add_interface_block(StorageClass storage, bool patch)
auto &fake_var = set<SPIRVariable>(var_id, ptr_type_id, storage);
set_decoration(var_id, DecorationLocation, input.first);
meta.strip_array = true;
meta.allow_local_declaration = false;
add_variable_to_interface_block(storage, ib_var_ref, ib_type, fake_var, meta);
}
}
@ -3462,8 +3462,7 @@ uint32_t CompilerMSL::add_interface_block(StorageClass storage, bool patch)
// The member indices were saved to the original variables, but after the members
// were sorted, those indices are now likely incorrect. Fix those up now.
if (!patch)
fix_up_interface_member_indices(storage, ib_type_id);
fix_up_interface_member_indices(storage, ib_type_id);
// For patch inputs, add one more member, holding the array of control point data.
if (get_execution_model() == ExecutionModelTessellationEvaluation && storage == StorageClassInput && patch &&
@ -6869,6 +6868,10 @@ bool CompilerMSL::emit_tessellation_access_chain(const uint32_t *ops, uint32_t l
bool flat_data = false;
bool ptr_is_chain = false;
bool multi_patch = get_execution_model() == ExecutionModelTessellationControl && msl_options.multi_patch_workgroup;
bool is_block = false;
if (var)
is_block = has_decoration(get_variable_data_type(*var).self, DecorationBlock);
if (var)
{
@ -6878,20 +6881,20 @@ bool CompilerMSL::emit_tessellation_access_chain(const uint32_t *ops, uint32_t l
flat_data = var->storage == StorageClassInput ||
(var->storage == StorageClassOutput && get_execution_model() == ExecutionModelTessellationControl);
if (patch && (!is_block || var->storage != StorageClassOutput))
flat_data = false;
// We might have a chained access chain, where
// we first take the access chain to the control point, and then we chain into a member or something similar.
// In this case, we need to skip gl_in/gl_out remapping.
// Also, skip ptr chain for patches.
ptr_is_chain = var->self != ID(ops[2]);
}
bool builtin_variable = false;
bool is_block = false;
bool variable_is_flat = false;
if (var)
is_block = has_decoration(get_variable_data_type(*var).self, DecorationBlock);
if (var && flat_data && !patch)
if (var && flat_data)
{
builtin_variable = is_builtin_variable(*var);
@ -6914,11 +6917,12 @@ bool CompilerMSL::emit_tessellation_access_chain(const uint32_t *ops, uint32_t l
bool masked = false;
if (is_block)
{
uint32_t relevant_member_index = patch ? 3 : 4;
// FIXME: This won't work properly if the application first access chains into gl_out element,
// then access chains into the member. Super weird, but theoretically possible ...
if (length >= 5)
if (length > relevant_member_index)
{
uint32_t mbr_idx = get<SPIRConstant>(ops[4]).scalar();
uint32_t mbr_idx = get<SPIRConstant>(ops[relevant_member_index]).scalar();
masked = is_stage_output_block_member_masked(*var, mbr_idx, true);
}
}
@ -6935,13 +6939,19 @@ bool CompilerMSL::emit_tessellation_access_chain(const uint32_t *ops, uint32_t l
indices.reserve(length - 3 + 1);
uint32_t first_non_array_index = ptr_is_chain ? 3 : 4;
VariableID stage_var_id = var->storage == StorageClassInput ? stage_in_ptr_var_id : stage_out_ptr_var_id;
uint32_t first_non_array_index = (ptr_is_chain ? 3 : 4) - (patch ? 1 : 0);
VariableID stage_var_id;
if (patch)
stage_var_id = patch_stage_out_var_id;
else
stage_var_id = var->storage == StorageClassInput ? stage_in_ptr_var_id : stage_out_ptr_var_id;
VariableID ptr = ptr_is_chain ? VariableID(ops[2]) : stage_var_id;
if (!ptr_is_chain)
if (!ptr_is_chain && !patch)
{
// Index into gl_in/gl_out with first array index.
indices.push_back(ops[3]);
indices.push_back(ops[first_non_array_index - 1]);
}
auto &result_ptr_type = get<SPIRType>(ops[0]);
@ -6956,11 +6966,12 @@ bool CompilerMSL::emit_tessellation_access_chain(const uint32_t *ops, uint32_t l
{
// Maybe this is a struct type in the input class, in which case
// we put it as a decoration on the corresponding member.
index = get_extended_member_decoration(var->self, get_constant(ops[first_non_array_index]).scalar(),
uint32_t mbr_idx = get_constant(ops[first_non_array_index]).scalar();
index = get_extended_member_decoration(var->self, mbr_idx,
SPIRVCrossDecorationInterfaceMemberIndex);
assert(index != uint32_t(-1));
i++;
type = &get<SPIRType>(type->member_types[get_constant(ops[first_non_array_index]).scalar()]);
type = &get<SPIRType>(type->member_types[mbr_idx]);
}
// In this case, we're poking into flattened structures and arrays, so now we have to
@ -7030,7 +7041,7 @@ bool CompilerMSL::emit_tessellation_access_chain(const uint32_t *ops, uint32_t l
if (!ptr_is_chain)
{
// This is the start of an access chain, use ptr_chain to index into control point array.
e = access_chain(ptr, indices.data(), uint32_t(indices.size()), result_ptr_type, &meta, true);
e = access_chain(ptr, indices.data(), uint32_t(indices.size()), result_ptr_type, &meta, !patch);
}
else
{

1
spirv_msl.hpp
View File

@ -803,6 +803,7 @@ protected:
};
std::unordered_map<uint32_t, LocationMeta> location_meta;
bool strip_array = false;
bool allow_local_declaration = false;
};
void add_variable_to_interface_block(spv::StorageClass storage, const std::string &ib_var_ref, SPIRType &ib_type,