blob: ea7f5fe6be3aba6120d87897419850b7bc4684a9 [file] [log] [blame]
// Copyright 2022 The Dawn & Tint Authors
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// 1. Redistributions of source code must retain the above copyright notice, this
// list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// 3. Neither the name of the copyright holder nor the names of its
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
// DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
// FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
// SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "src/tint/lang/wgsl/ast/transform/std140.h"
#include <algorithm>
#include <string>
#include <utility>
#include <variant>
#include "src/tint/lang/core/fluent_types.h"
#include "src/tint/lang/wgsl/program/clone_context.h"
#include "src/tint/lang/wgsl/program/program_builder.h"
#include "src/tint/lang/wgsl/resolver/resolve.h"
#include "src/tint/lang/wgsl/sem/index_accessor_expression.h"
#include "src/tint/lang/wgsl/sem/member_accessor_expression.h"
#include "src/tint/lang/wgsl/sem/module.h"
#include "src/tint/lang/wgsl/sem/struct.h"
#include "src/tint/lang/wgsl/sem/variable.h"
#include "src/tint/utils/containers/hashmap.h"
#include "src/tint/utils/containers/transform.h"
#include "src/tint/utils/macros/compiler.h"
#include "src/tint/utils/rtti/switch.h"
TINT_INSTANTIATE_TYPEINFO(tint::ast::transform::Std140);
using namespace tint::core::number_suffixes; // NOLINT
using namespace tint::core::fluent_types; // NOLINT
namespace tint::ast::transform {
namespace {
/// UniformVariable is used by Std140::State::AccessIndex to indicate the root uniform variable
struct UniformVariable {
/// @returns a hash code for this object
tint::HashCode HashCode() const { return 0; }
};
/// Inequality operator for UniformVariable
bool operator!=(const UniformVariable&, const UniformVariable&) {
return false;
}
/// DynamicIndex is used by Std140::State::AccessIndex to indicate a runtime-expression index
struct DynamicIndex {
size_t slot; // The index of the expression in Std140::State::AccessChain::dynamic_indices
/// @returns a hash code for this object
tint::HashCode HashCode() const { return Hash(slot); }
};
/// Inequality operator for DynamicIndex
bool operator!=(const DynamicIndex& a, const DynamicIndex& b) {
return a.slot != b.slot;
}
} // namespace
/// PIMPL state for the transform
struct Std140::State {
/// Constructor
/// @param program the source program
explicit State(const Program& program) : src(program) {}
/// Runs the transform
/// @returns the new program or SkipTransform if the transform is not required
ApplyResult Run() {
if (!ShouldRun()) {
// Transform is not required
return SkipTransform;
}
// Begin by creating forked types for any type that is used as a uniform buffer, that
// either directly or transitively contains a matrix that needs splitting for std140 layout.
ForkTypes();
// Next, replace all the uniform variables to use the forked types.
ReplaceUniformVarTypes();
// Finally, replace all expression chains that used the authored types with those that
// correctly use the forked types.
ctx.ReplaceAll([&](const Expression* expr) -> const Expression* {
if (auto access = AccessChainFor(expr)) {
if (!access->std140_mat_idx.has_value()) {
// loading a std140 type, which is not a whole or partial decomposed matrix
return LoadWithConvert(access.value());
}
if (!access->IsMatrixSubset() || // loading a whole matrix
std::holds_alternative<DynamicIndex>(
access->indices[*access->std140_mat_idx + 1])) {
// Whole object or matrix is loaded, or the matrix column is indexed with a
// non-constant index. Build a helper function to load the expression chain.
return LoadMatrixWithFn(access.value());
}
// Matrix column is statically indexed. Can be emitted as an inline expression.
return LoadSubMatrixInline(access.value());
}
// Expression isn't an access to a std140-layout uniform buffer.
// Just clone.
return nullptr;
});
ctx.Clone();
return resolver::Resolve(b);
}
/// @returns true if this transform should be run for the given program
bool ShouldRun() const {
// Returns true if the type needs to be forked for std140 usage.
auto needs_fork = [&](const core::type::Type* ty) {
while (auto* arr = ty->As<core::type::Array>()) {
ty = arr->ElemType();
}
if (auto* mat = ty->As<core::type::Matrix>()) {
if (MatrixNeedsDecomposing(mat)) {
return true;
}
}
return false;
};
// Scan structures for members that need forking
for (auto* ty : src.Types()) {
if (auto* str = ty->As<core::type::Struct>()) {
if (str->UsedAs(core::AddressSpace::kUniform)) {
for (auto* member : str->Members()) {
if (needs_fork(member->Type())) {
return true;
}
}
}
}
}
// Scan uniform variables that have types that need forking
for (auto* decl : src.AST().GlobalVariables()) {
auto* global = src.Sem().Get(decl);
if (global->AddressSpace() == core::AddressSpace::kUniform) {
if (needs_fork(global->Type()->UnwrapRef())) {
return true;
}
}
}
// If we reach here, no uniform variables use a type that needs forking for std140 layout
return false;
}
private:
/// Swizzle describes a vector swizzle
using Swizzle = tint::Vector<uint32_t, 4>;
/// AccessIndex describes a single access in an access chain.
/// The access is one of:
/// UniformVariable - the root uniform variable.
/// u32 - a static index on a struct, array index, matrix column or vector element.
/// DynamicIndex - a runtime index on an array, matrix column, or vector element.
/// Swizzle - a static vector swizzle.
using AccessIndex = std::variant<UniformVariable, u32, DynamicIndex, Swizzle>;
/// A vector of AccessIndex.
using AccessIndices = tint::Vector<AccessIndex, 8>;
/// A key used to cache load functions for an access chain.
struct LoadFnKey {
/// The root uniform buffer variable for the access chain.
const sem::GlobalVariable* var;
/// The chain of accesses indices.
AccessIndices indices;
/// @returns the hash code for the LoadFnKey
tint::HashCode HashCode() const { return Hash(var, indices); }
/// Equality operator
bool operator==(const LoadFnKey& other) const {
return var == other.var && indices == other.indices;
}
};
/// The source program
const Program& src;
/// The target program builder
ProgramBuilder b;
/// The clone context
program::CloneContext ctx = {&b, &src, /* auto_clone_symbols */ true};
/// Alias to the semantic info in src
const sem::Info& sem = src.Sem();
/// Alias to the symbols in src
const SymbolTable& sym = src.Symbols();
/// Map of load function signature, to the generated function
Hashmap<LoadFnKey, Symbol, 8> load_fns;
/// Map of std140-forked type to converter function name
Hashmap<const core::type::Type*, Symbol, 8> conv_fns;
// Uniform variables that have been modified to use a std140 type
Hashset<const sem::Variable*, 8> std140_uniforms;
// Map of original structure to 'std140' forked structure
Hashmap<const core::type::Struct*, Symbol, 8> std140_structs;
// Map of structure member in src of a matrix type, to list of decomposed column
// members in ctx.dst.
Hashmap<const core::type::StructMember*, tint::Vector<const StructMember*, 4>, 8>
std140_mat_members;
/// Describes a matrix that has been forked to a std140-structure holding the decomposed column
/// vectors of the matrix.
struct Std140Matrix {
/// The decomposed structure name (in ctx.dst)
Symbol name;
/// The column vector structure member names (in ctx.dst)
tint::Vector<Symbol, 4> columns;
};
// Map of matrix type in src, to decomposed column structure in ctx.dst.
Hashmap<const core::type::Matrix*, Std140Matrix, 8> std140_mats;
/// AccessChain describes a chain of access expressions to uniform buffer variable.
struct AccessChain {
/// The uniform buffer variable.
const sem::GlobalVariable* var;
/// The chain of access indices, starting with the first access on #var.
AccessIndices indices;
/// The runtime-evaluated expressions. This vector is indexed by the DynamicIndex::slot
tint::Vector<const sem::ValueExpression*, 8> dynamic_indices;
/// The type of the std140-decomposed matrix being accessed.
/// May be nullptr if the chain does not pass through a std140-decomposed matrix.
const core::type::Matrix* std140_mat_ty = nullptr;
/// The index in #indices of the access that resolves to the std140-decomposed matrix.
/// May hold no value if the chain does not pass through a std140-decomposed matrix.
std::optional<size_t> std140_mat_idx;
/// @returns true if the access chain is to part of (not the whole) std140-decomposed matrix
bool IsMatrixSubset() const {
return std140_mat_idx.has_value() && (std140_mat_idx.value() + 1 != indices.Length());
}
};
/// @returns true if the given matrix needs decomposing to column vectors for std140 layout.
/// Std140 layout require matrix stride to be 16, otherwise decomposing is needed.
static bool MatrixNeedsDecomposing(const core::type::Matrix* mat) {
return mat->ColumnStride() != 16;
}
/// ForkTypes walks the user-declared types in dependency order, forking structures that are
/// used as uniform buffers which (transitively) use matrices that need std140 decomposition to
/// column vectors. Populates the #std140_mat_members map, #std140_structs set and #std140_mats
/// map (via Std140Type()).
void ForkTypes() {
// For each module scope declaration...
for (auto* global : src.Sem().Module()->DependencyOrderedDeclarations()) {
// Check to see if this is a structure used by a uniform buffer...
auto* str = sem.Get<sem::Struct>(global);
if (str && str->UsedAs(core::AddressSpace::kUniform)) {
// Should this uniform buffer be forked for std140 usage?
bool fork_std140 = false;
tint::Vector<const StructMember*, 8> members;
for (auto* member : str->Members()) {
if (auto* mat = member->Type()->As<core::type::Matrix>()) {
// Is this member a matrix that needs decomposition for std140-layout?
if (MatrixNeedsDecomposing(mat)) {
// Structure member of matrix type needs decomposition.
fork_std140 = true;
// Replace the member with column vectors.
const auto name_prefix = PrefixForUniqueNames(
str->Declaration(), member->Name(), mat->Columns());
// Build a struct member for each column of the matrix
auto column_members = DecomposedMatrixStructMembers(
mat, name_prefix, member->Align(), member->Size());
// Add the member to the forked structure
for (auto* column_member : column_members) {
members.Push(column_member);
}
// Record that this matrix member was replaced with the N column
// members.
std140_mat_members.Add(member, std::move(column_members));
continue; // Next member
}
} else if (auto std140_ty = Std140Type(member->Type())) {
// Member is of a type that requires forking for std140-layout
fork_std140 = true;
auto attrs = ctx.Clone(member->Declaration()->attributes);
members.Push(b.Member(member->Name().Name(), std140_ty, std::move(attrs)));
continue; // Next member
}
// Nothing special about this member.
// Push the member in src to members without first cloning. We'll replace this
// with a cloned member once we know whether we need to fork the structure or
// not.
members.Push(member->Declaration());
}
// Did any of the members require forking the structure?
if (fork_std140) {
// Clone any members that have not already been cloned.
for (auto& member : members) {
if (member->generation_id == src.ID()) {
member = ctx.Clone(member);
}
}
// Create a new forked structure, and insert it just under the original
// structure.
auto name = b.Symbols().New(str->Name().Name() + "_std140");
auto* std140 = b.create<Struct>(b.Ident(name), std::move(members),
ctx.Clone(str->Declaration()->attributes));
ctx.InsertAfter(src.AST().GlobalDeclarations(), global, std140);
std140_structs.Add(str, name);
}
}
}
}
/// Walks the global variables, replacing the type of those that are a uniform buffer with a
/// type that has been forked for std140-layout.
/// Populates the #std140_uniforms set.
void ReplaceUniformVarTypes() {
for (auto* global : src.AST().GlobalVariables()) {
if (auto* var = global->As<Var>()) {
auto* v = sem.Get(var);
if (v->AddressSpace() == core::AddressSpace::kUniform) {
if (auto std140_ty = Std140Type(v->Type()->UnwrapRef())) {
ctx.Replace(global->type.expr, b.Expr(std140_ty));
std140_uniforms.Add(v);
}
}
}
}
}
/// @returns a unique structure member prefix for the splitting of a matrix member into @p count
/// column vector members. The new members must be suffixed with a zero-based index ranging from
/// `[0..count)`.
/// @param str the structure that will hold the uniquely named member.
/// @param unsuffixed the common name prefix to use for the new members.
/// @param count the number of members that need to be created.
std::string PrefixForUniqueNames(const Struct* str, Symbol unsuffixed, uint32_t count) const {
auto prefix = unsuffixed.Name();
// Keep on inserting '_' between the unsuffixed name and the suffix numbers until the name
// is unique.
while (true) {
prefix += "_";
Hashset<std::string, 4> strings;
for (uint32_t i = 0; i < count; i++) {
strings.Add(prefix + std::to_string(i));
}
bool unique = true;
for (auto* member : str->members) {
// The member name must be unique over the entire set of `count` suffixed names.
if (strings.Contains(member->name->symbol.Name())) {
unique = false;
break;
}
}
if (unique) {
return prefix;
}
}
}
/// @returns a new, forked std140 AST type for the corresponding non-forked semantic type.
/// If the semantic type is not split for std140-layout, then nullptr is returned.
/// @note will construct new std140 structures to hold decomposed matrices, populating
/// #std140_mats.
Type Std140Type(const core::type::Type* ty) {
return Switch(
ty, //
[&](const core::type::Struct* str) {
if (auto std140 = std140_structs.Get(str)) {
return b.ty(*std140);
}
return Type{};
},
[&](const core::type::Matrix* mat) {
if (MatrixNeedsDecomposing(mat)) {
auto std140_mat = std140_mats.GetOrAdd(mat, [&] {
auto name = b.Symbols().New("mat" + std::to_string(mat->Columns()) + "x" +
std::to_string(mat->Rows()) + "_" +
mat->Type()->FriendlyName());
auto members =
DecomposedMatrixStructMembers(mat, "col", mat->Align(), mat->Size());
b.Structure(name, members);
return Std140Matrix{
name,
tint::Transform(members,
[&](auto* member) { return member->name->symbol; }),
};
});
return b.ty(std140_mat.name);
}
return Type{};
},
[&](const core::type::Array* arr) {
if (auto std140 = Std140Type(arr->ElemType())) {
tint::Vector<const Attribute*, 1> attrs;
if (!arr->IsStrideImplicit()) {
attrs.Push(b.create<StrideAttribute>(arr->Stride()));
}
auto count = arr->ConstantCount();
if (TINT_UNLIKELY(!count)) {
// Non-constant counts should not be possible:
// * Override-expression counts can only be applied to workgroup arrays, and
// this method only handles types transitively used as uniform buffers.
// * Runtime-sized arrays cannot be used in uniform buffers.
TINT_ICE() << "unexpected non-constant array count";
}
return b.ty.array(std140, b.Expr(u32(count.value())), std::move(attrs));
}
return Type{};
});
}
/// @param mat the matrix to decompose (in src)
/// @param name_prefix the name prefix to apply to each of the returned column vector members.
/// @param align the alignment in bytes of the matrix.
/// @param size the size in bytes of the matrix.
/// @returns a vector of decomposed matrix column vectors as structure members (in ctx.dst).
tint::Vector<const StructMember*, 4> DecomposedMatrixStructMembers(
const core::type::Matrix* mat,
const std::string& name_prefix,
uint32_t align,
uint32_t size) {
// Replace the member with column vectors.
const auto num_columns = mat->Columns();
const auto column_size = mat->ColumnType()->Size();
const auto column_stride = mat->ColumnStride();
// Build a struct member for each column of the matrix
tint::Vector<const StructMember*, 4> out;
for (uint32_t i = 0; i < num_columns; i++) {
tint::Vector<const Attribute*, 1> attributes;
if ((i == 0) && mat->Align() != align) {
// The matrix was @align() annotated with a larger alignment
// than the natural alignment for the matrix. This extra padding
// needs to be applied to the first column vector.
attributes.Push(b.MemberAlign(i32(align)));
}
if ((i == num_columns - 1) &&
(column_stride * (num_columns - 1) + column_size) != size) {
// The matrix size is larger than the individual component vectors.
// This occurs with matNx3 matrices, as the last vec3 column has space for one extra
// trailing scalar, which is occupied by the matrix. It also applies to matrices
// with an explicit @size() attribute.
// Apply extra padding needs to the last column vector.
attributes.Push(
b.MemberSize(AInt(size - mat->ColumnType()->Align() * (num_columns - 1))));
}
// Build the member
const auto col_name = name_prefix + std::to_string(i);
const auto col_ty = CreateASTTypeFor(ctx, mat->ColumnType());
const auto* col_member = b.Member(col_name, col_ty, std::move(attributes));
// Record the member for std140_mat_members
out.Push(col_member);
}
return out;
}
/// Walks the @p ast_expr, constructing and returning an AccessChain.
/// @returns an AccessChain if the expression is an access to a std140-forked uniform buffer,
/// otherwise returns a std::nullopt.
std::optional<AccessChain> AccessChainFor(const Expression* ast_expr) {
auto* expr = sem.GetVal(ast_expr);
if (!expr) {
return std::nullopt;
}
AccessChain access;
// Start by looking at the root identifier. This must be a std140-forked uniform buffer.
access.var = tint::As<sem::GlobalVariable>(expr->RootIdentifier());
if (!access.var || !std140_uniforms.Contains(access.var)) {
// Not at std140-forked uniform buffer access chain.
return std::nullopt;
}
// Walk from the outer-most expression, inwards towards the root identifier.
while (true) {
enum class Action { kStop, kContinue, kError };
Action action = Switch(
expr->Unwrap(), //
[&](const sem::VariableUser* user) {
if (user->Variable() == access.var) {
// Walked all the way to the root identifier. We're done traversing.
access.indices.Push(UniformVariable{});
return Action::kStop;
}
if (TINT_LIKELY(user->Variable()->Type()->Is<core::type::Pointer>())) {
// Found a pointer. As the root identifier is a uniform buffer variable,
// this must be a pointer-let. Continue traversing from the let
// initializer.
expr = user->Variable()->Initializer();
return Action::kContinue;
}
TINT_ICE() << "unexpected variable found walking access chain: "
<< user->Variable()->Declaration()->name->symbol.Name();
},
[&](const sem::StructMemberAccess* a) {
// Is this a std140 decomposed matrix?
if (std140_mat_members.Contains(a->Member())) {
// Record this on the access.
access.std140_mat_idx = access.indices.Length();
access.std140_mat_ty = expr->Type()->UnwrapRef()->As<core::type::Matrix>();
}
// Structure member accesses are always statically indexed
access.indices.Push(u32(a->Member()->Index()));
expr = a->Object();
return Action::kContinue;
},
[&](const sem::IndexAccessorExpression* a) {
// Array, matrix or vector index.
if (auto* val = a->Index()->ConstantValue()) {
access.indices.Push(val->ValueAs<u32>());
} else {
access.indices.Push(DynamicIndex{access.dynamic_indices.Length()});
access.dynamic_indices.Push(a->Index());
}
expr = a->Object();
// Is the object a std140 decomposed matrix?
if (auto* mat = expr->Type()->UnwrapPtrOrRef()->As<core::type::Matrix>()) {
if (std140_mats.Contains(mat)) {
// Record this on the access.
access.std140_mat_idx = access.indices.Length();
access.std140_mat_ty = mat;
}
}
return Action::kContinue;
},
[&](const sem::Swizzle* s) {
// Vector swizzle.
if (s->Indices().Length() == 1) {
access.indices.Push(u32(s->Indices()[0]));
} else {
access.indices.Push(s->Indices());
}
expr = s->Object();
return Action::kContinue;
},
[&](const sem::ValueExpression* e) {
// Walk past indirection and address-of unary ops.
return Switch(e->Declaration(), //
[&](const UnaryOpExpression* u) {
switch (u->op) {
case core::UnaryOp::kAddressOf:
case core::UnaryOp::kIndirection:
expr = sem.GetVal(u->expr);
return Action::kContinue;
default:
TINT_ICE() << "unhandled unary op for access chain: "
<< u->op;
}
});
},
[&](Default) -> Action {
TINT_ICE() << "unhandled expression type for access chain\n"
<< "AST: " << expr->Declaration()->TypeInfo().name << "\n"
<< "SEM: " << expr->TypeInfo().name;
});
switch (action) {
case Action::kContinue:
continue;
case Action::kStop:
break;
case Action::kError:
return std::nullopt;
}
break;
}
// As the access walked from RHS to LHS, the last index operation applies to the source
// variable. We want this the other way around, so reverse the arrays and fix indicies.
std::reverse(access.indices.begin(), access.indices.end());
std::reverse(access.dynamic_indices.begin(), access.dynamic_indices.end());
if (access.std140_mat_idx.has_value()) {
access.std140_mat_idx = access.indices.Length() - *access.std140_mat_idx - 1;
}
for (auto& index : access.indices) {
if (auto* dyn_idx = std::get_if<DynamicIndex>(&index)) {
dyn_idx->slot = access.dynamic_indices.Length() - dyn_idx->slot - 1;
}
}
return access;
}
/// @returns a name suffix for a std140 -> non-std140 conversion function based on the type
/// being converted.
const std::string ConvertSuffix(const core::type::Type* ty) {
return Switch(
ty, //
[&](const core::type::Struct* str) { return str->Name().Name(); },
[&](const core::type::Array* arr) {
auto count = arr->ConstantCount();
if (TINT_UNLIKELY(!count)) {
// Non-constant counts should not be possible:
// * Override-expression counts can only be applied to workgroup arrays, and
// this method only handles types transitively used as uniform buffers.
// * Runtime-sized arrays cannot be used in uniform buffers.
TINT_ICE() << "unexpected non-constant array count";
}
return "arr" + std::to_string(count.value()) + "_" + ConvertSuffix(arr->ElemType());
},
[&](const core::type::Matrix* mat) {
return "mat" + std::to_string(mat->Columns()) + "x" + std::to_string(mat->Rows()) +
"_" + ConvertSuffix(mat->Type());
},
[&](const core::type::F32*) { return "f32"; }, //
[&](const core::type::F16*) { return "f16"; }, //
TINT_ICE_ON_NO_MATCH);
}
/// Generates and returns an expression that loads the value from a std140 uniform buffer,
/// converting the final result to a non-std140 type.
/// @param chain the access chain from a uniform buffer to the value to load.
const Expression* LoadWithConvert(const AccessChain& chain) {
const Expression* expr = nullptr;
const core::type::Type* ty = nullptr;
auto dynamic_index = [&](size_t idx) {
return ctx.Clone(chain.dynamic_indices[idx]->Declaration());
};
for (size_t i = 0; i < chain.indices.Length(); i++) {
auto [new_expr, new_ty, _] = BuildAccessExpr(expr, ty, chain, i, dynamic_index);
expr = new_expr;
ty = new_ty;
}
return Convert(ty, expr);
}
/// Generates and returns an expression that converts the expression @p expr of the
/// std140-forked type to the type @p ty. If @p expr is not a std140-forked type, then Convert()
/// will simply return @p expr.
/// @returns the converted value expression.
const Expression* Convert(const core::type::Type* ty, const Expression* expr) {
// Get an existing, or create a new function for converting the std140 type to ty.
auto fn = conv_fns.GetOrAdd(ty, [&] {
auto std140_ty = Std140Type(ty);
if (!std140_ty) {
// ty was not forked for std140.
return Symbol{};
}
// The converter function takes a single argument of the std140 type.
auto* param = b.Param("val", std140_ty);
tint::Vector<const Statement*, 3> stmts;
Switch(
ty, //
[&](const core::type::Struct* str) {
// Convert each of the structure members using either a converter function
// call, or by reassembling a std140 matrix from column vector members.
tint::Vector<const Expression*, 8> args;
for (auto* member : str->Members()) {
if (auto col_members = std140_mat_members.Get(member)) {
// std140 decomposed matrix. Reassemble.
auto mat_ty = CreateASTTypeFor(ctx, member->Type());
auto mat_args =
tint::Transform(*col_members, [&](const StructMember* m) {
return b.MemberAccessor(param, m->name->symbol);
});
args.Push(b.Call(mat_ty, std::move(mat_args)));
} else {
// Convert the member
args.Push(Convert(member->Type(),
b.MemberAccessor(param, member->Name().Name())));
}
}
stmts.Push(b.Return(b.Call(CreateASTTypeFor(ctx, ty), std::move(args))));
}, //
[&](const core::type::Matrix* mat) {
// Reassemble a std140 matrix from the structure of column vector members.
auto std140_mat = std140_mats.Get(mat);
if (TINT_LIKELY(std140_mat)) {
tint::Vector<const Expression*, 8> args;
// std140 decomposed matrix. Reassemble.
auto mat_ty = CreateASTTypeFor(ctx, mat);
auto mat_args = tint::Transform(std140_mat->columns, [&](Symbol name) {
return b.MemberAccessor(param, name);
});
stmts.Push(b.Return(b.Call(mat_ty, std::move(mat_args))));
} else {
TINT_ICE()
<< "failed to find std140 matrix info for: " << ty->FriendlyName();
}
}, //
[&](const core::type::Array* arr) {
// Converting an array. Create a function var for the converted array, and
// loop over the input elements, converting each and assigning the result to
// the local array.
auto* var = b.Var("arr", CreateASTTypeFor(ctx, ty));
auto* i = b.Var("i", b.ty.u32());
auto* dst_el = b.IndexAccessor(var, i);
auto* src_el = Convert(arr->ElemType(), b.IndexAccessor(param, i));
auto count = arr->ConstantCount();
if (TINT_UNLIKELY(!count)) {
// Non-constant counts should not be possible:
// * Override-expression counts can only be applied to workgroup arrays, and
// this method only handles types transitively used as uniform buffers.
// * Runtime-sized arrays cannot be used in uniform buffers.
TINT_ICE() << "unexpected non-constant array count";
}
stmts.Push(b.Decl(var));
stmts.Push(b.For(b.Decl(i), //
b.LessThan(i, u32(count.value())), //
b.Assign(i, b.Add(i, 1_a)), //
b.Block(b.Assign(dst_el, src_el))));
stmts.Push(b.Return(var));
}, //
TINT_ICE_ON_NO_MATCH);
// Generate the function
auto ret_ty = CreateASTTypeFor(ctx, ty);
auto fn_sym = b.Symbols().New("conv_" + ConvertSuffix(ty));
b.Func(fn_sym, tint::Vector{param}, ret_ty, std::move(stmts));
return fn_sym;
});
if (!fn.IsValid()) {
// Not a std140 type, nothing to convert.
return expr;
}
// Call the helper
return b.Call(fn, tint::Vector{expr});
}
/// Loads a part of, or a whole std140-decomposed matrix from a uniform buffer, using a helper
/// function which will be generated if it hasn't been already.
/// @param access the access chain from the uniform buffer to either the whole matrix or part of
/// the matrix (column, column-swizzle, or element).
/// @returns the loaded value expression.
const Expression* LoadMatrixWithFn(const AccessChain& access) {
// Get an existing, or create a new function for loading the uniform buffer value.
// This function is keyed off the uniform buffer variable and the access chain.
auto fn = load_fns.GetOrAdd(LoadFnKey{access.var, access.indices}, [&] {
if (access.IsMatrixSubset()) {
// Access chain passes through the matrix, but ends either at a column vector,
// column swizzle, or element.
return BuildLoadPartialMatrixFn(access);
}
// Access is to the whole matrix.
return BuildLoadWholeMatrixFn(access);
});
// Build the arguments
auto args = tint::Transform(access.dynamic_indices, [&](const sem::ValueExpression* e) {
return b.Call<u32>(ctx.Clone(e->Declaration()));
});
// Call the helper
return b.Call(fn, std::move(args));
}
/// Loads a part of a std140-decomposed matrix from a uniform buffer, inline (without calling a
/// helper function).
/// @param chain the access chain from the uniform buffer to part of the matrix (column,
/// column-swizzle, or element).
/// @note The matrix column must be statically indexed to use this method.
/// @returns the loaded value expression.
const Expression* LoadSubMatrixInline(const AccessChain& chain) {
// Method for generating dynamic index expressions.
// As this is inline, we can just clone the expression.
auto dynamic_index = [&](size_t idx) {
return ctx.Clone(chain.dynamic_indices[idx]->Declaration());
};
const Expression* expr = nullptr;
const core::type::Type* ty = nullptr;
// Build the expression up to, but not including the matrix member
auto std140_mat_idx = *chain.std140_mat_idx;
for (size_t i = 0; i < std140_mat_idx; i++) {
auto [new_expr, new_ty, _] = BuildAccessExpr(expr, ty, chain, i, dynamic_index);
expr = new_expr;
ty = new_ty;
}
// Access is to the std140 decomposed matrix.
// As this is accessing only part of the matrix, we just need to pick the right column
// vector member.
auto column_idx = std::get<u32>(chain.indices[std140_mat_idx + 1]);
if (auto* str = tint::As<core::type::Struct>(ty)) {
// Structure member matrix. The columns are decomposed into the structure.
auto mat_member_idx = std::get<u32>(chain.indices[std140_mat_idx]);
auto* mat_member = str->Members()[mat_member_idx];
auto mat_columns = *std140_mat_members.Get(mat_member);
expr = b.MemberAccessor(expr, mat_columns[column_idx]->name->symbol);
ty = mat_member->Type()->As<core::type::Matrix>()->ColumnType();
} else {
// Non-structure-member matrix. The columns are decomposed into a new, bespoke std140
// structure.
auto [new_expr, new_ty, _] =
BuildAccessExpr(expr, ty, chain, std140_mat_idx, dynamic_index);
expr = new_expr;
ty = new_ty;
auto* mat = ty->As<core::type::Matrix>();
auto std140_mat = std140_mats.Get(ty->As<core::type::Matrix>());
expr = b.MemberAccessor(expr, std140_mat->columns[column_idx]);
ty = mat->ColumnType();
}
// Build any remaining accesses into the column
for (size_t i = std140_mat_idx + 2; i < chain.indices.Length(); i++) {
auto [new_expr, new_ty, _] = BuildAccessExpr(expr, ty, chain, i, dynamic_index);
expr = new_expr;
ty = new_ty;
}
return expr;
}
/// Generates a function to load part of a std140-decomposed matrix from a uniform buffer.
/// The generated function will have a parameter per dynamic (runtime-evaluated) index in the
/// access chain.
/// The generated function uses a WGSL switch statement to dynamically select the decomposed
/// matrix column.
/// @param chain the access chain from the uniform buffer to part of the matrix (column,
/// column-swizzle, or element).
/// @note The matrix column must be dynamically indexed to use this method.
/// @returns the generated function name.
Symbol BuildLoadPartialMatrixFn(const AccessChain& chain) {
// Build the dynamic index parameters
auto dynamic_index_params = tint::Transform(chain.dynamic_indices, [&](auto*, size_t i) {
return b.Param("p" + std::to_string(i), b.ty.u32());
});
// Method for generating dynamic index expressions.
// These are passed in as arguments to the function.
auto dynamic_index = [&](size_t idx) {
return b.Expr(dynamic_index_params[idx]->name->symbol);
};
// Fetch the access chain indices of the matrix access and the parameter index that
// holds the matrix column index.
auto std140_mat_idx = *chain.std140_mat_idx;
auto column_param_idx = std::get<DynamicIndex>(chain.indices[std140_mat_idx + 1]).slot;
// Begin building the function name. This is extended with logic in the loop below
// (when column_idx == 0).
std::string name = "load";
// The switch cases
tint::Vector<const CaseStatement*, 4> cases;
// The function return type.
const core::type::Type* ret_ty = nullptr;
// Build switch() cases for each column of the matrix
auto num_columns = chain.std140_mat_ty->Columns();
for (uint32_t column_idx = 0; column_idx < num_columns; column_idx++) {
const Expression* expr = nullptr;
const core::type::Type* ty = nullptr;
// Build the expression up to, but not including the matrix
for (size_t i = 0; i < std140_mat_idx; i++) {
auto [new_expr, new_ty, access_name] =
BuildAccessExpr(expr, ty, chain, i, dynamic_index);
expr = new_expr;
ty = new_ty;
if (column_idx == 0) {
name += "_" + access_name;
}
}
if (auto* str = tint::As<core::type::Struct>(ty)) {
// Structure member matrix. The columns are decomposed into the structure.
auto mat_member_idx = std::get<u32>(chain.indices[std140_mat_idx]);
auto* mat_member = str->Members()[mat_member_idx];
if (column_idx == 0) {
name +=
"_" + mat_member->Name().Name() + "_p" + std::to_string(column_param_idx);
}
auto mat_columns = *std140_mat_members.Get(mat_member);
expr = b.MemberAccessor(expr, mat_columns[column_idx]->name->symbol);
ty = mat_member->Type()->As<core::type::Matrix>()->ColumnType();
} else {
// Non-structure-member matrix. The columns are decomposed into a new, bespoke
// std140 structure.
auto [new_expr, new_ty, mat_name] =
BuildAccessExpr(expr, ty, chain, std140_mat_idx, dynamic_index);
expr = new_expr;
ty = new_ty;
if (column_idx == 0) {
name += "_" + mat_name + "_p" + std::to_string(column_param_idx);
}
auto* mat = ty->As<core::type::Matrix>();
auto std140_mat = std140_mats.Get(ty->As<core::type::Matrix>());
expr = b.MemberAccessor(expr, std140_mat->columns[column_idx]);
ty = mat->ColumnType();
}
// Build the rest of the expression, skipping over the column index.
for (size_t i = std140_mat_idx + 2; i < chain.indices.Length(); i++) {
auto [new_expr, new_ty, access_name] =
BuildAccessExpr(expr, ty, chain, i, dynamic_index);
expr = new_expr;
ty = new_ty;
if (column_idx == 0) {
name += "_" + access_name;
}
}
if (column_idx == 0) {
ret_ty = ty;
}
auto* case_sel = b.CaseSelector(b.Expr(u32(column_idx)));
auto* case_body = b.Block(tint::Vector{b.Return(expr)});
cases.Push(b.Case(case_sel, case_body));
}
// Build the default case (required in WGSL).
// This just returns a zero value of the return type, as the index must be out of
// bounds.
cases.Push(b.DefaultCase(b.Block(b.Return(b.Call(CreateASTTypeFor(ctx, ret_ty))))));
auto* column_selector = dynamic_index(column_param_idx);
auto* stmt = b.Switch(column_selector, std::move(cases));
auto fn_sym = b.Symbols().New(name);
b.Func(fn_sym, std::move(dynamic_index_params), CreateASTTypeFor(ctx, ret_ty),
tint::Vector{stmt});
return fn_sym;
}
/// Generates a function to load a whole std140-decomposed matrix from a uniform buffer.
/// The generated function will have a parameter per dynamic (runtime-evaluated) index in the
/// access chain.
/// @param chain the access chain from the uniform buffer to the whole std140-decomposed
/// matrix.
/// @returns the generated function name.
Symbol BuildLoadWholeMatrixFn(const AccessChain& chain) {
// Build the dynamic index parameters
auto dynamic_index_params = tint::Transform(chain.dynamic_indices, [&](auto*, size_t i) {
return b.Param("p" + std::to_string(i), b.ty.u32());
});
// Method for generating dynamic index expressions.
// These are passed in as arguments to the function.
auto dynamic_index = [&](size_t idx) {
return b.Expr(dynamic_index_params[idx]->name->symbol);
};
const Expression* expr = nullptr;
const core::type::Type* ty = nullptr;
std::string name = "load";
// Build the expression up to, but not including the matrix member
auto std140_mat_idx = *chain.std140_mat_idx;
for (size_t i = 0; i < std140_mat_idx; i++) {
auto [new_expr, new_ty, access_name] =
BuildAccessExpr(expr, ty, chain, i, dynamic_index);
expr = new_expr;
ty = new_ty;
name += "_" + access_name;
}
tint::Vector<const Statement*, 2> stmts;
// Create a temporary pointer to the structure that holds the matrix columns
auto* let = b.Let("s", b.AddressOf(expr));
stmts.Push(b.Decl(let));
tint::Vector<const MemberAccessorExpression*, 4> columns;
if (auto* str = tint::As<core::type::Struct>(ty)) {
// Structure member matrix. The columns are decomposed into the structure.
auto mat_member_idx = std::get<u32>(chain.indices[std140_mat_idx]);
auto* mat_member = str->Members()[mat_member_idx];
auto mat_columns = *std140_mat_members.Get(mat_member);
columns = tint::Transform(mat_columns, [&](auto* column_member) {
return b.MemberAccessor(b.Deref(let), column_member->name->symbol);
});
ty = mat_member->Type();
name += "_" + mat_member->Name().Name();
} else {
// Non-structure-member matrix. The columns are decomposed into a new, bespoke
// std140 structure.
auto [new_expr, new_ty, mat_name] =
BuildAccessExpr(expr, ty, chain, std140_mat_idx, dynamic_index);
expr = new_expr;
auto* mat = ty->As<core::type::Matrix>();
auto std140_mat = std140_mats.Get(ty->As<core::type::Matrix>());
columns = tint::Transform(std140_mat->columns, [&](auto column_name) {
return b.MemberAccessor(b.Deref(let), column_name);
});
ty = mat;
name += "_" + mat_name;
}
// Reconstruct the matrix from the columns
expr = b.Call(CreateASTTypeFor(ctx, chain.std140_mat_ty), std::move(columns));
// Have the function return the constructed matrix
stmts.Push(b.Return(expr));
// Build the function
auto ret_ty = CreateASTTypeFor(ctx, ty);
auto fn_sym = b.Symbols().New(name);
b.Func(fn_sym, std::move(dynamic_index_params), ret_ty, std::move(stmts));
return fn_sym;
}
/// Return type of BuildAccessExpr()
struct ExprTypeName {
/// The new, post-access expression
const Expression* expr;
/// The type of #expr
const core::type::Type* type;
/// A name segment which can be used to build sensible names for helper functions
std::string name;
};
/// Builds a single access in an access chain.
/// @param lhs the expression to index using @p access
/// @param ty the type of the expression @p lhs
/// @param chain the access index to perform on @p lhs
/// @param dynamic_index a function that obtains the i'th dynamic index
/// @returns a ExprTypeName which holds the new expression, new type and a name segment which
/// can be used for creating helper function names.
ExprTypeName BuildAccessExpr(const Expression* lhs,
const core::type::Type* ty,
const AccessChain& chain,
size_t index,
std::function<const Expression*(size_t)> dynamic_index) {
auto& access = chain.indices[index];
if (std::get_if<UniformVariable>(&access)) {
const auto symbol = chain.var->Declaration()->name->symbol;
const auto* expr = b.Expr(ctx.Clone(symbol));
const auto name = symbol.Name();
ty = chain.var->Type()->UnwrapRef();
return {expr, ty, name};
}
if (auto* dyn_idx = std::get_if<DynamicIndex>(&access)) {
/// The access uses a dynamic (runtime-expression) index.
auto name = "p" + std::to_string(dyn_idx->slot);
return Switch(
ty, //
[&](const core::type::Array* arr) -> ExprTypeName {
auto* idx = dynamic_index(dyn_idx->slot);
auto* expr = b.IndexAccessor(lhs, idx);
return {expr, arr->ElemType(), name};
}, //
[&](const core::type::Matrix* mat) -> ExprTypeName {
auto* idx = dynamic_index(dyn_idx->slot);
auto* expr = b.IndexAccessor(lhs, idx);
return {expr, mat->ColumnType(), name};
}, //
[&](const core::type::Vector* vec) -> ExprTypeName {
auto* idx = dynamic_index(dyn_idx->slot);
auto* expr = b.IndexAccessor(lhs, idx);
return {expr, vec->Type(), name};
}, //
TINT_ICE_ON_NO_MATCH);
}
if (auto* swizzle = std::get_if<Swizzle>(&access)) {
/// The access is a vector swizzle.
return Switch(
ty, //
[&](const core::type::Vector* vec) -> ExprTypeName {
static const char xyzw[] = {'x', 'y', 'z', 'w'};
std::string rhs;
for (auto el : *swizzle) {
rhs += xyzw[el];
}
auto swizzle_ty = src.Types().Find<core::type::Vector>(
vec->Type(), static_cast<uint32_t>(swizzle->Length()));
auto* expr = b.MemberAccessor(lhs, rhs);
return {expr, swizzle_ty, rhs};
}, //
TINT_ICE_ON_NO_MATCH);
}
/// The access is a static index.
auto idx = std::get<u32>(access);
return Switch(
ty, //
[&](const core::type::Struct* str) -> ExprTypeName {
auto* member = str->Members()[idx];
auto member_name = member->Name().Name();
auto* expr = b.MemberAccessor(lhs, member_name);
ty = member->Type();
return {expr, ty, member_name};
}, //
[&](const core::type::Array* arr) -> ExprTypeName {
auto* expr = b.IndexAccessor(lhs, idx);
return {expr, arr->ElemType(), std::to_string(idx)};
}, //
[&](const core::type::Matrix* mat) -> ExprTypeName {
auto* expr = b.IndexAccessor(lhs, idx);
return {expr, mat->ColumnType(), std::to_string(idx)};
}, //
[&](const core::type::Vector* vec) -> ExprTypeName {
auto* expr = b.IndexAccessor(lhs, idx);
return {expr, vec->Type(), std::to_string(idx)};
}, //
TINT_ICE_ON_NO_MATCH);
}
};
Std140::Std140() = default;
Std140::~Std140() = default;
Transform::ApplyResult Std140::Apply(const Program& src, const DataMap&, DataMap&) const {
return State(src).Run();
}
} // namespace tint::ast::transform