3
0
Fork 0
forked from suyu/suyu

shader: Implement HSET2

This commit is contained in:
FernandoS27 2021-03-21 03:48:40 +01:00 committed by ameerj
parent ed6cd3c94a
commit 9e213fd861
5 changed files with 119 additions and 14 deletions

View file

@ -82,6 +82,7 @@ add_library(shader_recompiler STATIC
frontend/maxwell/translate/impl/half_floating_point_helper.cpp frontend/maxwell/translate/impl/half_floating_point_helper.cpp
frontend/maxwell/translate/impl/half_floating_point_helper.h frontend/maxwell/translate/impl/half_floating_point_helper.h
frontend/maxwell/translate/impl/half_floating_point_multiply.cpp frontend/maxwell/translate/impl/half_floating_point_multiply.cpp
frontend/maxwell/translate/impl/half_floating_point_set.cpp
frontend/maxwell/translate/impl/impl.cpp frontend/maxwell/translate/impl/impl.cpp
frontend/maxwell/translate/impl/impl.h frontend/maxwell/translate/impl/impl.h
frontend/maxwell/translate/impl/integer_add.cpp frontend/maxwell/translate/impl/integer_add.cpp

View file

@ -105,8 +105,8 @@ INST(HMUL2_cbuf, "HMUL2 (cbuf)", "0111 100- 1--- ----")
INST(HMUL2_imm, "HMUL2 (imm)", "0111 100- 0--- ----") INST(HMUL2_imm, "HMUL2 (imm)", "0111 100- 0--- ----")
INST(HMUL2_32I, "HMUL2_32I", "0010 101- ---- ----") INST(HMUL2_32I, "HMUL2_32I", "0010 101- ---- ----")
INST(HSET2_reg, "HSET2 (reg)", "0101 1101 0001 1---") INST(HSET2_reg, "HSET2 (reg)", "0101 1101 0001 1---")
INST(HSET2_cbuf, "HSET2 (cbuf)", "0111 1100 1--- ----") INST(HSET2_cbuf, "HSET2 (cbuf)", "0111 110- 1--- ----")
INST(HSET2_imm, "HSET2 (imm)", "0111 1100 0--- ----") INST(HSET2_imm, "HSET2 (imm)", "0111 110- 0--- ----")
INST(HSETP2_reg, "HSETP2 (reg)", "0101 1101 0010 0---") INST(HSETP2_reg, "HSETP2 (reg)", "0101 1101 0010 0---")
INST(HSETP2_cbuf, "HSETP2 (cbuf)", "0111 111- 1--- ----") INST(HSETP2_cbuf, "HSETP2 (cbuf)", "0111 111- 1--- ----")
INST(HSETP2_imm, "HSETP2 (imm)", "0111 111- 0--- ----") INST(HSETP2_imm, "HSETP2 (imm)", "0111 111- 0--- ----")

View file

@ -5,6 +5,7 @@
#include "common/common_types.h" #include "common/common_types.h"
#include "shader_recompiler/exception.h" #include "shader_recompiler/exception.h"
#include "shader_recompiler/frontend/maxwell/translate/impl/common_encoding.h" #include "shader_recompiler/frontend/maxwell/translate/impl/common_encoding.h"
#include "shader_recompiler/frontend/maxwell/translate/impl/common_funcs.h"
#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h" #include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
namespace Shader::Maxwell { namespace Shader::Maxwell {

View file

@ -0,0 +1,115 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_helper.h"
namespace Shader::Maxwell {
namespace {
void HSET2(TranslatorVisitor& v, u64 insn, const IR::U32& src_b, bool bf, bool ftz, bool neg_b,
bool abs_b, FPCompareOp compare_op, Swizzle swizzle_b) {
union {
u64 insn;
BitField<0, 8, IR::Reg> dest_reg;
BitField<8, 8, IR::Reg> src_a_reg;
BitField<39, 3, IR::Pred> pred;
BitField<42, 1, u64> neg_pred;
BitField<43, 1, u64> neg_a;
BitField<45, 2, BooleanOp> bop;
BitField<44, 1, u64> abs_a;
BitField<47, 2, Swizzle> swizzle_a;
} const hset2{insn};
auto [lhs_a, rhs_a]{Extract(v.ir, v.X(hset2.src_a_reg), hset2.swizzle_a)};
auto [lhs_b, rhs_b]{Extract(v.ir, src_b, swizzle_b)};
// TODO: Implement FP16 FloatingPointCompare
//if (lhs_a.Type() != lhs_b.Type()) {
if (lhs_a.Type() == IR::Type::F16) {
lhs_a = v.ir.FPConvert(32, lhs_a);
rhs_a = v.ir.FPConvert(32, rhs_a);
}
if (lhs_b.Type() == IR::Type::F16) {
lhs_b = v.ir.FPConvert(32, lhs_b);
rhs_b = v.ir.FPConvert(32, rhs_b);
}
//}
lhs_a = v.ir.FPAbsNeg(lhs_a, hset2.abs_a != 0, hset2.neg_a != 0);
rhs_a = v.ir.FPAbsNeg(rhs_a, hset2.abs_a != 0, hset2.neg_a != 0);
lhs_b = v.ir.FPAbsNeg(lhs_b, abs_b, neg_b);
rhs_b = v.ir.FPAbsNeg(rhs_b, abs_b, neg_b);
const IR::FpControl control{
.no_contraction{false},
.rounding{IR::FpRounding::DontCare},
.fmz_mode{ftz ? IR::FmzMode::FTZ : IR::FmzMode::None},
};
IR::U1 pred{v.ir.GetPred(hset2.pred)};
if (hset2.neg_pred != 0) {
pred = v.ir.LogicalNot(pred);
}
const IR::U1 cmp_result_lhs{FloatingPointCompare(v.ir, lhs_a, lhs_b, compare_op, control)};
const IR::U1 cmp_result_rhs{FloatingPointCompare(v.ir, rhs_a, rhs_b, compare_op, control)};
const IR::U1 bop_result_lhs{PredicateCombine(v.ir, cmp_result_lhs, pred, hset2.bop)};
const IR::U1 bop_result_rhs{PredicateCombine(v.ir, cmp_result_rhs, pred, hset2.bop)};
const u32 true_value = bf ? 0x3c00 : 0xffff;
const IR::U32 true_val_lhs{v.ir.Imm32(true_value)};
const IR::U32 true_val_rhs{v.ir.Imm32(true_value << 16)};
const IR::U32 fail_result{v.ir.Imm32(0)};
const IR::U32 result_lhs{v.ir.Select(bop_result_lhs, true_val_lhs, fail_result)};
const IR::U32 result_rhs{v.ir.Select(bop_result_rhs, true_val_rhs, fail_result)};
v.X(hset2.dest_reg, IR::U32{v.ir.BitwiseOr(result_lhs, result_rhs)});
}
} // Anonymous namespace
void TranslatorVisitor::HSET2_reg(u64 insn) {
union {
u64 insn;
BitField<30, 1, u64> abs_b;
BitField<49, 1, u64> bf;
BitField<31, 1, u64> neg_b;
BitField<50, 1, u64> ftz;
BitField<35, 4, FPCompareOp> compare_op;
BitField<28, 2, Swizzle> swizzle_b;
} const hset2{insn};
HSET2(*this, insn, GetReg20(insn), hset2.bf != 0, hset2.ftz != 0, hset2.neg_b != 0,
hset2.abs_b != 0, hset2.compare_op, hset2.swizzle_b);
}
void TranslatorVisitor::HSET2_cbuf(u64 insn) {
union {
u64 insn;
BitField<53, 1, u64> bf;
BitField<56, 1, u64> neg_b;
BitField<54, 1, u64> ftz;
BitField<49, 4, FPCompareOp> compare_op;
} const hset2{insn};
HSET2(*this, insn, GetCbuf(insn), hset2.bf != 0, hset2.ftz != 0, hset2.neg_b != 0, false,
hset2.compare_op, Swizzle::F32);
}
void TranslatorVisitor::HSET2_imm(u64 insn) {
union {
u64 insn;
BitField<53, 1, u64> bf;
BitField<54, 1, u64> ftz;
BitField<49, 4, FPCompareOp> compare_op;
BitField<56, 1, u64> neg_high;
BitField<30, 9, u64> high;
BitField<29, 1, u64> neg_low;
BitField<20, 9, u64> low;
} const hset2{insn};
const u32 imm{static_cast<u32>(hset2.low << 6) | ((hset2.neg_low != 0 ? 1 : 0) << 15) |
static_cast<u32>(hset2.high << 22) | ((hset2.neg_high != 0 ? 1 : 0) << 31)};
HSET2(*this, insn, ir.Imm32(imm), hset2.bf != 0, hset2.ftz != 0, false, false,
hset2.compare_op, Swizzle::H1_H0);
}
} // namespace Shader::Maxwell

View file

@ -181,18 +181,6 @@ void TranslatorVisitor::GETLMEMBASE(u64) {
ThrowNotImplemented(Opcode::GETLMEMBASE); ThrowNotImplemented(Opcode::GETLMEMBASE);
} }
void TranslatorVisitor::HSET2_reg(u64) {
ThrowNotImplemented(Opcode::HSET2_reg);
}
void TranslatorVisitor::HSET2_cbuf(u64) {
ThrowNotImplemented(Opcode::HSET2_cbuf);
}
void TranslatorVisitor::HSET2_imm(u64) {
ThrowNotImplemented(Opcode::HSET2_imm);
}
void TranslatorVisitor::HSETP2_reg(u64) { void TranslatorVisitor::HSETP2_reg(u64) {
ThrowNotImplemented(Opcode::HSETP2_reg); ThrowNotImplemented(Opcode::HSETP2_reg);
} }