Shaders: Implemented multiple-word loads and stores to and from attribute memory.

This seems to be an optimization performed by nouveau.
This commit is contained in:
Subv 2018-09-08 16:13:02 -05:00
parent df5a44a40b
commit c878a819d7
2 changed files with 58 additions and 7 deletions

View File

@ -67,6 +67,13 @@ private:
u64 value{};
};
enum class AttributeSize : u64 {
Word = 0,
DoubleWord = 1,
TripleWord = 2,
QuadWord = 3,
};
union Attribute {
Attribute() = default;
@ -87,9 +94,10 @@ union Attribute {
};
union {
BitField<20, 10, u64> immediate;
BitField<22, 2, u64> element;
BitField<24, 6, Index> index;
BitField<47, 3, u64> size;
BitField<47, 3, AttributeSize> size;
} fmt20;
union {

View File

@ -1772,13 +1772,34 @@ private:
case OpCode::Type::Memory: {
switch (opcode->GetId()) {
case OpCode::Id::LD_A: {
ASSERT_MSG(instr.attribute.fmt20.size == 0, "untested");
// Note: Shouldn't this be interp mode flat? As in no interpolation made.
ASSERT_MSG(instr.gpr8.Value() == Register::ZeroIndex,
"Indirect attribute loads are not supported");
ASSERT_MSG((instr.attribute.fmt20.immediate.Value() % sizeof(u32)) == 0,
"Unaligned attribute loads are not supported");
Tegra::Shader::IpaMode input_mode{Tegra::Shader::IpaInterpMode::Perspective,
Tegra::Shader::IpaSampleMode::Default};
regs.SetRegisterToInputAttibute(instr.gpr0, instr.attribute.fmt20.element,
instr.attribute.fmt20.index, input_mode);
u32 next_element = instr.attribute.fmt20.element;
u32 next_index = static_cast<u32>(instr.attribute.fmt20.index.Value());
const auto LoadNextElement = [&](u32 reg_offset) {
regs.SetRegisterToInputAttibute(instr.gpr0.Value() + reg_offset, next_element,
static_cast<Attribute::Index>(next_index),
input_mode);
// Load the next attribute element into the following register. If the element
// to load goes beyond the vec4 size, load the first element of the next
// attribute.
next_element = (next_element + 1) % 4;
next_index = next_index + (next_element == 0 ? 1 : 0);
};
const u32 num_words = static_cast<u32>(instr.attribute.fmt20.size.Value()) + 1;
for (u32 reg_offset = 0; reg_offset < num_words; ++reg_offset) {
LoadNextElement(reg_offset);
}
break;
}
case OpCode::Id::LD_C: {
@ -1820,9 +1841,31 @@ private:
break;
}
case OpCode::Id::ST_A: {
ASSERT_MSG(instr.attribute.fmt20.size == 0, "untested");
regs.SetOutputAttributeToRegister(instr.attribute.fmt20.index,
instr.attribute.fmt20.element, instr.gpr0);
ASSERT_MSG(instr.gpr8.Value() == Register::ZeroIndex,
"Indirect attribute loads are not supported");
ASSERT_MSG((instr.attribute.fmt20.immediate.Value() % sizeof(u32)) == 0,
"Unaligned attribute loads are not supported");
u32 next_element = instr.attribute.fmt20.element;
u32 next_index = static_cast<u32>(instr.attribute.fmt20.index.Value());
const auto StoreNextElement = [&](u32 reg_offset) {
regs.SetOutputAttributeToRegister(static_cast<Attribute::Index>(next_index),
next_element,
instr.gpr0.Value() + reg_offset);
// Load the next attribute element into the following register. If the element
// to load goes beyond the vec4 size, load the first element of the next
// attribute.
next_element = (next_element + 1) % 4;
next_index = next_index + (next_element == 0 ? 1 : 0);
};
const u32 num_words = static_cast<u32>(instr.attribute.fmt20.size.Value()) + 1;
for (u32 reg_offset = 0; reg_offset < num_words; ++reg_offset) {
StoreNextElement(reg_offset);
}
break;
}
case OpCode::Id::TEX: {