1
1
Fork 0
forked from suyu/suyu

Merge pull request #2812 from ReinUsesLisp/f2i-selector

shader_ir/conversion: Implement F2I and F2F F16 selector
This commit is contained in:
bunnei 2019-09-03 22:35:33 -04:00 committed by GitHub
commit 81fbc5370d
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 23 additions and 7 deletions

View file

@ -1020,7 +1020,6 @@ union Instruction {
} iset; } iset;
union { union {
BitField<41, 2, u64> selector; // i2i and i2f only
BitField<45, 1, u64> negate_a; BitField<45, 1, u64> negate_a;
BitField<49, 1, u64> abs_a; BitField<49, 1, u64> abs_a;
BitField<10, 2, Register::Size> src_size; BitField<10, 2, Register::Size> src_size;
@ -1046,6 +1045,13 @@ union Instruction {
} }
} f2f; } f2f;
union {
BitField<41, 2, u64> selector;
} int_src;
union {
BitField<41, 1, u64> selector;
} float_src;
} conversion; } conversion;
union { union {

View file

@ -14,6 +14,12 @@ using Tegra::Shader::Instruction;
using Tegra::Shader::OpCode; using Tegra::Shader::OpCode;
using Tegra::Shader::Register; using Tegra::Shader::Register;
namespace {
constexpr OperationCode GetFloatSelector(u64 selector) {
return selector == 0 ? OperationCode::FCastHalf0 : OperationCode::FCastHalf1;
}
} // Anonymous namespace
u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) { u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
const Instruction instr = {program_code[pc]}; const Instruction instr = {program_code[pc]};
const auto opcode = OpCode::Decode(instr); const auto opcode = OpCode::Decode(instr);
@ -22,7 +28,7 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
case OpCode::Id::I2I_R: case OpCode::Id::I2I_R:
case OpCode::Id::I2I_C: case OpCode::Id::I2I_C:
case OpCode::Id::I2I_IMM: { case OpCode::Id::I2I_IMM: {
UNIMPLEMENTED_IF(instr.conversion.selector); UNIMPLEMENTED_IF(instr.conversion.int_src.selector != 0);
UNIMPLEMENTED_IF(instr.conversion.dst_size != Register::Size::Word); UNIMPLEMENTED_IF(instr.conversion.dst_size != Register::Size::Word);
UNIMPLEMENTED_IF(instr.alu.saturate_d); UNIMPLEMENTED_IF(instr.alu.saturate_d);
@ -57,8 +63,8 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
case OpCode::Id::I2F_R: case OpCode::Id::I2F_R:
case OpCode::Id::I2F_C: case OpCode::Id::I2F_C:
case OpCode::Id::I2F_IMM: { case OpCode::Id::I2F_IMM: {
UNIMPLEMENTED_IF(instr.conversion.int_src.selector != 0);
UNIMPLEMENTED_IF(instr.conversion.dst_size == Register::Size::Long); UNIMPLEMENTED_IF(instr.conversion.dst_size == Register::Size::Long);
UNIMPLEMENTED_IF(instr.conversion.selector);
UNIMPLEMENTED_IF_MSG(instr.generates_cc, UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in I2F is not implemented"); "Condition codes generation in I2F is not implemented");
@ -113,8 +119,10 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
}(); }();
if (instr.conversion.src_size == Register::Size::Short) { if (instr.conversion.src_size == Register::Size::Short) {
// TODO: figure where extract is sey in the encoding value = Operation(GetFloatSelector(instr.conversion.float_src.selector), NO_PRECISE,
value = Operation(OperationCode::FCastHalf0, PRECISE, value); std::move(value));
} else {
ASSERT(instr.conversion.float_src.selector == 0);
} }
value = GetOperandAbsNegFloat(value, instr.conversion.abs_a, instr.conversion.negate_a); value = GetOperandAbsNegFloat(value, instr.conversion.abs_a, instr.conversion.negate_a);
@ -169,8 +177,10 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
}(); }();
if (instr.conversion.src_size == Register::Size::Short) { if (instr.conversion.src_size == Register::Size::Short) {
// TODO: figure where extract is sey in the encoding value = Operation(GetFloatSelector(instr.conversion.float_src.selector), NO_PRECISE,
value = Operation(OperationCode::FCastHalf0, PRECISE, value); std::move(value));
} else {
ASSERT(instr.conversion.float_src.selector == 0);
} }
value = GetOperandAbsNegFloat(value, instr.conversion.abs_a, instr.conversion.negate_a); value = GetOperandAbsNegFloat(value, instr.conversion.abs_a, instr.conversion.negate_a);