shader: Implement DMNMX, DSET, DSETP

src/shader_recompiler/CMakeLists.txt

@@ -65,8 +65,11 @@ add_library(shader_recompiler STATIC
     frontend/maxwell/translate/impl/common_funcs.h
     frontend/maxwell/translate/impl/condition_code_set.cpp
     frontend/maxwell/translate/impl/double_add.cpp
+    frontend/maxwell/translate/impl/double_compare_and_set.cpp
     frontend/maxwell/translate/impl/double_fused_multiply_add.cpp
+    frontend/maxwell/translate/impl/double_min_max.cpp
     frontend/maxwell/translate/impl/double_multiply.cpp
+    frontend/maxwell/translate/impl/double_set_predicate.cpp
     frontend/maxwell/translate/impl/exit_program.cpp
     frontend/maxwell/translate/impl/find_leading_one.cpp
     frontend/maxwell/translate/impl/floating_point_add.cpp

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -152,24 +152,7 @@ void DefineEntryPoint(Environment& env, EmitContext& ctx, Id main) {
 
 void SetupDenormControl(const Profile& profile, const IR::Program& program, EmitContext& ctx,
                         Id main_func) {
-    if (!profile.support_float_controls) {
-        return;
-    }
     const Info& info{program.info};
-    if (!info.uses_fp32_denorms_flush && !info.uses_fp32_denorms_preserve &&
-        !info.uses_fp16_denorms_flush && !info.uses_fp16_denorms_preserve) {
-        return;
-    }
-    ctx.AddExtension("SPV_KHR_float_controls");
-
-    if (info.uses_fp16 && profile.support_fp16_signed_zero_nan_preserve) {
-        ctx.AddCapability(spv::Capability::SignedZeroInfNanPreserve);
-        ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve, 16U);
-    }
-    if (profile.support_fp32_signed_zero_nan_preserve) {
-        ctx.AddCapability(spv::Capability::SignedZeroInfNanPreserve);
-        ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve, 32U);
-    }
     if (info.uses_fp32_denorms_flush && info.uses_fp32_denorms_preserve) {
         // LOG_ERROR(HW_GPU, "Fp32 denorm flush and preserve on the same shader");
     } else if (info.uses_fp32_denorms_flush) {
@@ -210,6 +193,22 @@ void SetupDenormControl(const Profile& profile, const IR::Program& program, Emit
     }
 }
 
+void SetupSignedNanCapabilities(const Profile& profile, const IR::Program& program,
+                                EmitContext& ctx, Id main_func) {
+    if (program.info.uses_fp16 && profile.support_fp16_signed_zero_nan_preserve) {
+        ctx.AddCapability(spv::Capability::SignedZeroInfNanPreserve);
+        ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve, 16U);
+    }
+    if (profile.support_fp32_signed_zero_nan_preserve) {
+        ctx.AddCapability(spv::Capability::SignedZeroInfNanPreserve);
+        ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve, 32U);
+    }
+    if (program.info.uses_fp64 && profile.support_fp64_signed_zero_nan_preserve) {
+        ctx.AddCapability(spv::Capability::SignedZeroInfNanPreserve);
+        ctx.AddExecutionMode(main_func, spv::ExecutionMode::SignedZeroInfNanPreserve, 64U);
+    }
+}
+
 void SetupCapabilities(const Profile& profile, const Info& info, EmitContext& ctx) {
     if (info.uses_sampled_1d) {
         ctx.AddCapability(spv::Capability::Sampled1D);
@@ -260,7 +259,11 @@ std::vector<u32> EmitSPIRV(const Profile& profile, Environment& env, IR::Program
     EmitContext ctx{profile, program, binding};
     const Id main{DefineMain(ctx, program)};
     DefineEntryPoint(env, ctx, main);
-    SetupDenormControl(profile, program, ctx, main);
+    if (profile.support_float_controls) {
+        ctx.AddExtension("SPV_KHR_float_controls");
+        SetupDenormControl(profile, program, ctx, main);
+        SetupSignedNanCapabilities(profile, program, ctx, main);
+    }
     SetupCapabilities(profile, program.info, ctx);
     return ctx.Assemble();
 }

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -136,6 +136,7 @@ Id EmitSelectU32(EmitContext& ctx, Id cond, Id true_value, Id false_value);
 Id EmitSelectU64(EmitContext& ctx, Id cond, Id true_value, Id false_value);
 Id EmitSelectF16(EmitContext& ctx, Id cond, Id true_value, Id false_value);
 Id EmitSelectF32(EmitContext& ctx, Id cond, Id true_value, Id false_value);
+Id EmitSelectF64(EmitContext& ctx, Id cond, Id true_value, Id false_value);
 void EmitBitCastU16F16(EmitContext& ctx);
 Id EmitBitCastU32F32(EmitContext& ctx, Id value);
 void EmitBitCastU64F64(EmitContext& ctx);

src/shader_recompiler/backend/spirv/emit_spirv_select.cpp

@@ -35,4 +35,8 @@ Id EmitSelectF32(EmitContext& ctx, Id cond, Id true_value, Id false_value) {
     return ctx.OpSelect(ctx.F32[1], cond, true_value, false_value);
 }
 
+Id EmitSelectF64(EmitContext& ctx, Id cond, Id true_value, Id false_value) {
+    return ctx.OpSelect(ctx.F64[1], cond, true_value, false_value);
+}
+
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/frontend/ir/ir_emitter.cpp

@@ -529,6 +529,8 @@ Value IREmitter::Select(const U1& condition, const Value& true_value, const Valu
         return Inst(Opcode::SelectU64, condition, true_value, false_value);
     case Type::F32:
         return Inst(Opcode::SelectF32, condition, true_value, false_value);
+    case Type::F64:
+        return Inst(Opcode::SelectF64, condition, true_value, false_value);
     default:
         throw InvalidArgument("Invalid type {}", true_value.Type());
     }

src/shader_recompiler/frontend/ir/opcodes.inc

@@ -131,6 +131,7 @@ OPCODE(SelectU32,                                           U32,            U1,
 OPCODE(SelectU64,                                           U64,            U1,             U64,            U64,                                            )
 OPCODE(SelectF16,                                           F16,            U1,             F16,            F16,                                            )
 OPCODE(SelectF32,                                           F32,            U1,             F32,            F32,                                            )
+OPCODE(SelectF64,                                           F64,            U1,             F64,            F64,                                            )
 
 // Bitwise conversions
 OPCODE(BitCastU16F16,                                       U16,            F16,                                                                            )

src/shader_recompiler/frontend/maxwell/maxwell.inc

@@ -37,8 +37,8 @@ INST(DFMA_reg,     "DFMA (reg)",     "0101 1011 0111 ----")
 INST(DFMA_rc,      "DFMA (rc)",      "0101 0011 0111 ----")
 INST(DFMA_cr,      "DFMA (cr)",      "0100 1011 0111 ----")
 INST(DFMA_imm,     "DFMA (imm)",     "0011 011- 0111 ----")
-INST(DMNMX_reg,    "DMNMX (reg)",    "0100 1100 0101 0---")
+INST(DMNMX_reg,    "DMNMX (reg)",    "0101 1100 0101 0---")
-INST(DMNMX_cbuf,   "DMNMX (cbuf)",   "0101 1100 0101 0---")
+INST(DMNMX_cbuf,   "DMNMX (cbuf)",   "0100 1100 0101 0---")
 INST(DMNMX_imm,    "DMNMX (imm)",    "0011 100- 0101 0---")
 INST(DMUL_reg,     "DMUL (reg)",     "0101 1100 1000 0---")
 INST(DMUL_cbuf,    "DMUL (cbuf)",    "0100 1100 1000 0---")

src/shader_recompiler/frontend/maxwell/translate/impl/double_compare_and_set.cpp

@@ -0,0 +1,59 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "shader_recompiler/frontend/maxwell/translate/impl/common_funcs.h"
+#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
+
+namespace Shader::Maxwell {
+namespace {
+void DSET(TranslatorVisitor& v, u64 insn, const IR::F64& src_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> negate_a;
+        BitField<44, 1, u64> abs_b;
+        BitField<45, 2, BooleanOp> bop;
+        BitField<48, 4, FPCompareOp> compare_op;
+        BitField<52, 1, u64> bf;
+        BitField<53, 1, u64> negate_b;
+        BitField<54, 1, u64> abs_a;
+    } const dset{insn};
+
+    const IR::F64 op_a{v.ir.FPAbsNeg(v.D(dset.src_a_reg), dset.abs_a != 0, dset.negate_a != 0)};
+    const IR::F64 op_b{v.ir.FPAbsNeg(src_b, dset.abs_b != 0, dset.negate_b != 0)};
+
+    IR::U1 pred{v.ir.GetPred(dset.pred)};
+    if (dset.neg_pred != 0) {
+        pred = v.ir.LogicalNot(pred);
+    }
+    const IR::U1 cmp_result{FloatingPointCompare(v.ir, op_a, op_b, dset.compare_op)};
+    const IR::U1 bop_result{PredicateCombine(v.ir, cmp_result, pred, dset.bop)};
+
+    const IR::U32 one_mask{v.ir.Imm32(-1)};
+    const IR::U32 fp_one{v.ir.Imm32(0x3f800000)};
+    const IR::U32 fail_result{v.ir.Imm32(0)};
+    const IR::U32 pass_result{dset.bf == 0 ? one_mask : fp_one};
+
+    v.X(dset.dest_reg, IR::U32{v.ir.Select(bop_result, pass_result, fail_result)});
+}
+} // Anonymous namespace
+
+void TranslatorVisitor::DSET_reg(u64 insn) {
+    DSET(*this, insn, GetDoubleReg20(insn));
+}
+
+void TranslatorVisitor::DSET_cbuf(u64 insn) {
+    DSET(*this, insn, GetDoubleCbuf(insn));
+}
+
+void TranslatorVisitor::DSET_imm(u64 insn) {
+    DSET(*this, insn, GetDoubleImm20(insn));
+}
+
+} // namespace Shader::Maxwell

src/shader_recompiler/frontend/maxwell/translate/impl/double_min_max.cpp

@@ -0,0 +1,50 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
+
+namespace Shader::Maxwell {
+namespace {
+void DMNMX(TranslatorVisitor& v, u64 insn, const IR::F64& src_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<45, 1, u64> negate_b;
+        BitField<46, 1, u64> abs_a;
+        BitField<48, 1, u64> negate_a;
+        BitField<49, 1, u64> abs_b;
+    } const dmnmx{insn};
+
+    const IR::U1 pred{v.ir.GetPred(dmnmx.pred)};
+    const IR::F64 op_a{v.ir.FPAbsNeg(v.D(dmnmx.src_a_reg), dmnmx.abs_a != 0, dmnmx.negate_a != 0)};
+    const IR::F64 op_b{v.ir.FPAbsNeg(src_b, dmnmx.abs_b != 0, dmnmx.negate_b != 0)};
+
+    IR::F64 max{v.ir.FPMax(op_a, op_b)};
+    IR::F64 min{v.ir.FPMin(op_a, op_b)};
+
+    if (dmnmx.neg_pred != 0) {
+        std::swap(min, max);
+    }
+    v.D(dmnmx.dest_reg, IR::F64{v.ir.Select(pred, min, max)});
+}
+} // Anonymous namespace
+
+void TranslatorVisitor::DMNMX_reg(u64 insn) {
+    DMNMX(*this, insn, GetDoubleReg20(insn));
+}
+
+void TranslatorVisitor::DMNMX_cbuf(u64 insn) {
+    DMNMX(*this, insn, GetDoubleCbuf(insn));
+}
+
+void TranslatorVisitor::DMNMX_imm(u64 insn) {
+    DMNMX(*this, insn, GetDoubleImm20(insn));
+}
+
+} // namespace Shader::Maxwell

src/shader_recompiler/frontend/maxwell/translate/impl/double_set_predicate.cpp

@@ -0,0 +1,54 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "shader_recompiler/frontend/maxwell/translate/impl/common_funcs.h"
+#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
+
+namespace Shader::Maxwell {
+namespace {
+void DSETP(TranslatorVisitor& v, u64 insn, const IR::F64& src_b) {
+    union {
+        u64 insn;
+        BitField<0, 3, IR::Pred> dest_pred_b;
+        BitField<3, 3, IR::Pred> dest_pred_a;
+        BitField<6, 1, u64> negate_b;
+        BitField<7, 1, u64> abs_a;
+        BitField<8, 8, IR::Reg> src_a_reg;
+        BitField<39, 3, IR::Pred> bop_pred;
+        BitField<42, 1, u64> neg_bop_pred;
+        BitField<43, 1, u64> negate_a;
+        BitField<44, 1, u64> abs_b;
+        BitField<45, 2, BooleanOp> bop;
+        BitField<48, 4, FPCompareOp> compare_op;
+    } const dsetp{insn};
+
+    const IR::F64 op_a{v.ir.FPAbsNeg(v.D(dsetp.src_a_reg), dsetp.abs_a != 0, dsetp.negate_a != 0)};
+    const IR::F64 op_b{v.ir.FPAbsNeg(src_b, dsetp.abs_b != 0, dsetp.negate_b != 0)};
+
+    const BooleanOp bop{dsetp.bop};
+    const FPCompareOp compare_op{dsetp.compare_op};
+    const IR::U1 comparison{FloatingPointCompare(v.ir, op_a, op_b, compare_op)};
+    const IR::U1 bop_pred{v.ir.GetPred(dsetp.bop_pred, dsetp.neg_bop_pred != 0)};
+    const IR::U1 result_a{PredicateCombine(v.ir, comparison, bop_pred, bop)};
+    const IR::U1 result_b{PredicateCombine(v.ir, v.ir.LogicalNot(comparison), bop_pred, bop)};
+    v.ir.SetPred(dsetp.dest_pred_a, result_a);
+    v.ir.SetPred(dsetp.dest_pred_b, result_b);
+}
+} // Anonymous namespace
+
+void TranslatorVisitor::DSETP_reg(u64 insn) {
+    DSETP(*this, insn, GetDoubleReg20(insn));
+}
+
+void TranslatorVisitor::DSETP_cbuf(u64 insn) {
+    DSETP(*this, insn, GetDoubleCbuf(insn));
+}
+
+void TranslatorVisitor::DSETP_imm(u64 insn) {
+    DSETP(*this, insn, GetDoubleImm20(insn));
+}
+
+} // namespace Shader::Maxwell

src/shader_recompiler/frontend/maxwell/translate/impl/floating_point_min_max.cpp

@@ -24,7 +24,7 @@ void FMNMX(TranslatorVisitor& v, u64 insn, const IR::F32& src_b) {
 
     const IR::U1 pred{v.ir.GetPred(fmnmx.pred)};
     const IR::F32 op_a{v.ir.FPAbsNeg(v.F(fmnmx.src_a_reg), fmnmx.abs_a != 0, fmnmx.negate_a != 0)};
-    const IR::F32 op_b = v.ir.FPAbsNeg(src_b, fmnmx.abs_b != 0, fmnmx.negate_b != 0);
+    const IR::F32 op_b{v.ir.FPAbsNeg(src_b, fmnmx.abs_b != 0, fmnmx.negate_b != 0)};
 
     const IR::FpControl control{
         .no_contraction{false},

src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp

@@ -81,42 +81,6 @@ void TranslatorVisitor::DEPBAR() {
     // DEPBAR is a no-op
 }
 
-void TranslatorVisitor::DMNMX_reg(u64) {
-    ThrowNotImplemented(Opcode::DMNMX_reg);
-}
-
-void TranslatorVisitor::DMNMX_cbuf(u64) {
-    ThrowNotImplemented(Opcode::DMNMX_cbuf);
-}
-
-void TranslatorVisitor::DMNMX_imm(u64) {
-    ThrowNotImplemented(Opcode::DMNMX_imm);
-}
-
-void TranslatorVisitor::DSET_reg(u64) {
-    ThrowNotImplemented(Opcode::DSET_reg);
-}
-
-void TranslatorVisitor::DSET_cbuf(u64) {
-    ThrowNotImplemented(Opcode::DSET_cbuf);
-}
-
-void TranslatorVisitor::DSET_imm(u64) {
-    ThrowNotImplemented(Opcode::DSET_imm);
-}
-
-void TranslatorVisitor::DSETP_reg(u64) {
-    ThrowNotImplemented(Opcode::DSETP_reg);
-}
-
-void TranslatorVisitor::DSETP_cbuf(u64) {
-    ThrowNotImplemented(Opcode::DSETP_cbuf);
-}
-
-void TranslatorVisitor::DSETP_imm(u64) {
-    ThrowNotImplemented(Opcode::DSETP_imm);
-}
-
 void TranslatorVisitor::FCHK_reg(u64) {
     ThrowNotImplemented(Opcode::FCHK_reg);
 }

src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp

@@ -130,6 +130,7 @@ void VisitUsages(Info& info, IR::Inst& inst) {
     case IR::Opcode::CompositeInsertF64x2:
     case IR::Opcode::CompositeInsertF64x3:
     case IR::Opcode::CompositeInsertF64x4:
+    case IR::Opcode::SelectF64:
     case IR::Opcode::BitCastU64F64:
     case IR::Opcode::BitCastF64U64:
     case IR::Opcode::PackDouble2x32:

src/shader_recompiler/ir_opt/constant_propagation_pass.cpp

@@ -229,7 +229,6 @@ void FoldISub32(IR::Inst& inst) {
     }
 }
 
-template <typename T>
 void FoldSelect(IR::Inst& inst) {
     const IR::Value cond{inst.Arg(0)};
     if (cond.IsImmediate()) {
@@ -340,8 +339,15 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
         return FoldBitCast<IR::Opcode::BitCastU32F32, u32, f32>(inst, IR::Opcode::BitCastF32U32);
     case IR::Opcode::IAdd64:
         return FoldAdd<u64>(block, inst);
+    case IR::Opcode::SelectU1:
+    case IR::Opcode::SelectU8:
+    case IR::Opcode::SelectU16:
     case IR::Opcode::SelectU32:
-        return FoldSelect<u32>(inst);
+    case IR::Opcode::SelectU64:
+    case IR::Opcode::SelectF16:
+    case IR::Opcode::SelectF32:
+    case IR::Opcode::SelectF64:
+        return FoldSelect(inst);
     case IR::Opcode::LogicalAnd:
         return FoldLogicalAnd(inst);
     case IR::Opcode::LogicalOr:

src/shader_recompiler/profile.h

@@ -18,6 +18,7 @@ struct Profile {
     bool support_fp32_denorm_flush{};
     bool support_fp16_signed_zero_nan_preserve{};
     bool support_fp32_signed_zero_nan_preserve{};
+    bool support_fp64_signed_zero_nan_preserve{};
 
     // FClamp is broken and OpFMax + OpFMin should be used instead
     bool has_broken_spirv_clamp{};

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -244,6 +244,8 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::GPU& gpu_,
             float_control.shaderSignedZeroInfNanPreserveFloat16 != VK_FALSE,
         .support_fp32_signed_zero_nan_preserve =
             float_control.shaderSignedZeroInfNanPreserveFloat32 != VK_FALSE,
+        .support_fp64_signed_zero_nan_preserve =
+            float_control.shaderSignedZeroInfNanPreserveFloat64 != VK_FALSE,
         .has_broken_spirv_clamp = driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS_KHR,
     };
 }