shader: Implement HFMA2

2021-03-21 01:03:21 +01:00 · 2021-03-21 01:03:21 +01:00 · 28dff6a629
parent 76c8a962ac
commit 28dff6a629
5 changed files with 192 additions and 20 deletions
--- a/src/shader_recompiler/CMakeLists.txt
+++ b/src/shader_recompiler/CMakeLists.txt
@ -78,6 +78,7 @@ add_library(shader_recompiler STATIC
    frontend/maxwell/translate/impl/floating_point_range_reduction.cpp
    frontend/maxwell/translate/impl/floating_point_set_predicate.cpp
    frontend/maxwell/translate/impl/half_floating_point_add.cpp
+    frontend/maxwell/translate/impl/half_floating_point_fused_multiply_add.cpp
    frontend/maxwell/translate/impl/half_floating_point_helper.cpp
    frontend/maxwell/translate/impl/half_floating_point_helper.h
    frontend/maxwell/translate/impl/impl.cpp
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_fused_multiply_add.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_fused_multiply_add.cpp
@ -0,0 +1,170 @@
+// 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 HFMA2(TranslatorVisitor& v, u64 insn, Merge merge, Swizzle swizzle_a, bool neg_b, bool neg_c,
+           Swizzle swizzle_b, Swizzle swizzle_c, const IR::U32& src_b, const IR::U32& src_c,
+           bool sat, HalfPrecision precision) {
+    union {
+        u64 raw;
+        BitField<0, 8, IR::Reg> dest_reg;
+        BitField<8, 8, IR::Reg> src_a;
+    } const hfma2{insn};
+
+    auto [lhs_a, rhs_a]{Extract(v.ir, v.X(hfma2.src_a), swizzle_a)};
+    auto [lhs_b, rhs_b]{Extract(v.ir, src_b, swizzle_b)};
+    auto [lhs_c, rhs_c]{Extract(v.ir, src_c, swizzle_c)};
+    const bool promotion{lhs_a.Type() != lhs_b.Type() || lhs_a.Type() != lhs_c.Type()};
+    if (promotion) {
+        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);
+        }
+        if (lhs_c.Type() == IR::Type::F16) {
+            lhs_c = v.ir.FPConvert(32, lhs_c);
+            rhs_c = v.ir.FPConvert(32, rhs_c);
+        }
+    }
+
+    lhs_b = v.ir.FPAbsNeg(lhs_b, false, neg_b);
+    rhs_b = v.ir.FPAbsNeg(rhs_b, false, neg_b);
+
+    lhs_c = v.ir.FPAbsNeg(lhs_c, false, neg_c);
+    rhs_c = v.ir.FPAbsNeg(rhs_c, false, neg_c);
+
+    const IR::FpControl fp_control{
+        .no_contraction{true},
+        .rounding{IR::FpRounding::DontCare},
+        .fmz_mode{HalfPrecision2FmzMode(precision)},
+    };
+    IR::F16F32F64 lhs{v.ir.FPFma(lhs_a, lhs_b, lhs_c, fp_control)};
+    IR::F16F32F64 rhs{v.ir.FPFma(rhs_a, rhs_b, rhs_c, fp_control)};
+    if (precision == HalfPrecision::FMZ && !sat) {
+        // Do not implement FMZ if SAT is enabled, as it does the logic for us.
+        // On D3D9 mode, anything * 0 is zero, even NAN and infinity
+        const IR::F32 zero{v.ir.Imm32(0.0f)};
+        const IR::U1 lhs_zero_a{v.ir.FPEqual(lhs_a, zero)};
+        const IR::U1 lhs_zero_b{v.ir.FPEqual(lhs_b, zero)};
+        const IR::U1 lhs_any_zero{v.ir.LogicalOr(lhs_zero_a, lhs_zero_b)};
+        lhs = IR::F16F32F64{v.ir.Select(lhs_any_zero, lhs_c, lhs)};
+
+        const IR::U1 rhs_zero_a{v.ir.FPEqual(rhs_a, zero)};
+        const IR::U1 rhs_zero_b{v.ir.FPEqual(rhs_b, zero)};
+        const IR::U1 rhs_any_zero{v.ir.LogicalOr(rhs_zero_a, rhs_zero_b)};
+        rhs = IR::F16F32F64{v.ir.Select(rhs_any_zero, rhs_c, rhs)};
+    }
+    if (sat) {
+        lhs = v.ir.FPSaturate(lhs);
+        rhs = v.ir.FPSaturate(rhs);
+    }
+    if (promotion) {
+        lhs = v.ir.FPConvert(16, lhs);
+        rhs = v.ir.FPConvert(16, rhs);
+    }
+    v.X(hfma2.dest_reg, MergeResult(v.ir, hfma2.dest_reg, lhs, rhs, merge));
+}
+
+void HFMA2(TranslatorVisitor& v, u64 insn, bool neg_b, bool neg_c, Swizzle swizzle_b,
+           Swizzle swizzle_c, const IR::U32& src_b, const IR::U32& src_c, bool sat,
+           HalfPrecision precision) {
+    union {
+        u64 raw;
+        BitField<47, 2, Swizzle> swizzle_a;
+        BitField<49, 2, Merge> merge;
+    } const hfma2{insn};
+
+    HFMA2(v, insn, hfma2.merge, hfma2.swizzle_a, neg_b, neg_c, swizzle_b, swizzle_c, src_b, src_c,
+          sat, precision);
+}
+
+} // namespace
+
+void TranslatorVisitor::HFMA2_reg(u64 insn) {
+    union {
+        u64 raw;
+        BitField<28, 2, Swizzle> swizzle_b;
+        BitField<32, 1, u64> saturate;
+        BitField<31, 1, u64> neg_b;
+        BitField<30, 1, u64> neg_c;
+        BitField<35, 2, Swizzle> swizzle_c;
+        BitField<37, 2, HalfPrecision> precision;
+    } const hfma2{insn};
+
+    HFMA2(*this, insn, hfma2.neg_b != 0, hfma2.neg_c != 0, hfma2.swizzle_b, hfma2.swizzle_c,
+          GetReg20(insn), GetReg39(insn), hfma2.saturate != 0, hfma2.precision);
+}
+
+void TranslatorVisitor::HFMA2_rc(u64 insn) {
+    union {
+        u64 raw;
+        BitField<51, 1, u64> neg_c;
+        BitField<52, 1, u64> saturate;
+        BitField<53, 2, Swizzle> swizzle_b;
+        BitField<56, 1, u64> neg_b;
+        BitField<57, 2, HalfPrecision> precision;
+    } const hfma2{insn};
+
+    HFMA2(*this, insn, hfma2.neg_b != 0, hfma2.neg_c != 0, hfma2.swizzle_b, Swizzle::F32,
+          GetReg39(insn), GetCbuf(insn), hfma2.saturate != 0, hfma2.precision);
+}
+
+void TranslatorVisitor::HFMA2_cr(u64 insn) {
+    union {
+        u64 raw;
+        BitField<51, 1, u64> neg_c;
+        BitField<52, 1, u64> saturate;
+        BitField<53, 2, Swizzle> swizzle_c;
+        BitField<56, 1, u64> neg_b;
+        BitField<57, 2, HalfPrecision> precision;
+    } const hfma2{insn};
+
+    HFMA2(*this, insn, hfma2.neg_b != 0, hfma2.neg_c != 0, Swizzle::F32, hfma2.swizzle_c,
+          GetCbuf(insn), GetReg39(insn), hfma2.saturate != 0, hfma2.precision);
+}
+
+void TranslatorVisitor::HFMA2_imm(u64 insn) {
+    union {
+        u64 raw;
+        BitField<51, 1, u64> neg_c;
+        BitField<52, 1, u64> saturate;
+        BitField<53, 2, Swizzle> swizzle_c;
+
+        BitField<56, 1, u64> neg_high;
+        BitField<30, 9, u64> high;
+        BitField<29, 1, u64> neg_low;
+        BitField<20, 9, u64> low;
+        BitField<57, 2, HalfPrecision> precision;
+    } const hfma2{insn};
+
+    const u32 imm{static_cast<u32>(hfma2.low << 6) | ((hfma2.neg_low != 0 ? 1 : 0) << 15) |
+                  static_cast<u32>(hfma2.high << 22) | ((hfma2.neg_high != 0 ? 1 : 0) << 31)};
+
+    HFMA2(*this, insn, false, hfma2.neg_c != 0, Swizzle::H1_H0, hfma2.swizzle_c, ir.Imm32(imm),
+          GetReg39(insn), hfma2.saturate != 0, hfma2.precision);
+}
+
+void TranslatorVisitor::HFMA2_32I(u64 insn) {
+    union {
+        u64 raw;
+        BitField<0, 8, IR::Reg> src_c;
+        BitField<20, 32, u64> imm32;
+        BitField<52, 1, u64> neg_c;
+        BitField<53, 2, Swizzle> swizzle_a;
+        BitField<55, 2, HalfPrecision> precision;
+    } const hfma2{insn};
+
+    const u32 imm{static_cast<u32>(hfma2.imm32)};
+    HFMA2(*this, insn, Merge::H1_H0, hfma2.swizzle_a, false, hfma2.neg_c != 0, Swizzle::H1_H0,
+          Swizzle::H1_H0, ir.Imm32(imm), X(hfma2.src_c), false, hfma2.precision);
+}
+
+} // namespace Shader::Maxwell
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_helper.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_helper.cpp
@ -6,6 +6,19 @@

 namespace Shader::Maxwell {

+IR::FmzMode HalfPrecision2FmzMode(HalfPrecision precision) {
+    switch (precision) {
+    case HalfPrecision::None:
+        return IR::FmzMode::None;
+    case HalfPrecision::FTZ:
+        return IR::FmzMode::FTZ;
+    case HalfPrecision::FMZ:
+        return IR::FmzMode::FMZ;
+    default:
+        return IR::FmzMode::DontCare;
+    }
+}
+
 std::pair<IR::F16F32F64, IR::F16F32F64> Extract(IR::IREmitter& ir, IR::U32 value, Swizzle swizzle) {
    switch (swizzle) {
    case Swizzle::H1_H0: {
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_helper.h
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_helper.h
@ -23,6 +23,14 @@ enum class Swizzle : u64 {
    H1_H1,
 };

+enum class HalfPrecision : u64 {
+    None = 0,
+    FTZ = 1,
+    FMZ = 2,
+};
+
+IR::FmzMode HalfPrecision2FmzMode(HalfPrecision precision);
+
 std::pair<IR::F16F32F64, IR::F16F32F64> Extract(IR::IREmitter& ir, IR::U32 value, Swizzle swizzle);

 IR::U32 MergeResult(IR::IREmitter& ir, IR::Reg dest, const IR::F16& lhs, const IR::F16& rhs,
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp
@ -181,26 +181,6 @@ void TranslatorVisitor::GETLMEMBASE(u64) {
    ThrowNotImplemented(Opcode::GETLMEMBASE);
 }

-void TranslatorVisitor::HFMA2_reg(u64) {
-    ThrowNotImplemented(Opcode::HFMA2_reg);
-}
-
-void TranslatorVisitor::HFMA2_rc(u64) {
-    ThrowNotImplemented(Opcode::HFMA2_rc);
-}
-
-void TranslatorVisitor::HFMA2_cr(u64) {
-    ThrowNotImplemented(Opcode::HFMA2_cr);
-}
-
-void TranslatorVisitor::HFMA2_imm(u64) {
-    ThrowNotImplemented(Opcode::HFMA2_imm);
-}
-
-void TranslatorVisitor::HFMA2_32I(u64) {
-    ThrowNotImplemented(Opcode::HFMA2_32I);
-}
-
 void TranslatorVisitor::HMUL2_reg(u64) {
    ThrowNotImplemented(Opcode::HMUL2_reg);
 }