From dc1a9a3bed2aa9b0851f07976b0c687172aa3edc Mon Sep 17 00:00:00 2001
From: FernandoS27 <fsahmkow27@gmail.com>
Date: Fri, 26 Mar 2021 20:51:05 +0100
Subject: shader: Implement TLD
---
src/shader_recompiler/frontend/ir/opcodes.inc | 6 +-
src/shader_recompiler/frontend/maxwell/maxwell.inc | 4 +-
.../maxwell/translate/impl/not_implemented.cpp | 8 -
.../maxwell/translate/impl/texture_load.cpp | 165 +++++++++++++++++++++
4 files changed, 170 insertions(+), 13 deletions(-)
create mode 100644 src/shader_recompiler/frontend/maxwell/translate/impl/texture_load.cpp
(limited to 'src/shader_recompiler/frontend')
diff --git a/src/shader_recompiler/frontend/ir/opcodes.inc b/src/shader_recompiler/frontend/ir/opcodes.inc
index 717aa71caa..302b8471df 100644
--- a/src/shader_recompiler/frontend/ir/opcodes.inc
+++ b/src/shader_recompiler/frontend/ir/opcodes.inc
@@ -378,7 +378,7 @@ OPCODE(BindlessImageSampleDrefImplicitLod, F32, U32,
OPCODE(BindlessImageSampleDrefExplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
OPCODE(BindlessImageGather, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(BindlessImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, )
-OPCODE(BindlessImageFetch, F32x4, U32, Opaque, U32, U32, )
+OPCODE(BindlessImageFetch, F32x4, U32, Opaque, Opaque, Opaque, Opaque, )
OPCODE(BindlessImageQueryDimensions, U32x4, U32, U32, )
OPCODE(BoundImageSampleImplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
@@ -387,7 +387,7 @@ OPCODE(BoundImageSampleDrefImplicitLod, F32, U32,
OPCODE(BoundImageSampleDrefExplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
OPCODE(BoundImageGather, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(BoundImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, )
-OPCODE(BoundImageFetch, F32x4, U32, Opaque, U32, U32, )
+OPCODE(BoundImageFetch, F32x4, U32, Opaque, Opaque, Opaque, Opaque, )
OPCODE(BoundImageQueryDimensions, U32x4, U32, U32, )
OPCODE(ImageSampleImplicitLod, F32x4, U32, Opaque, Opaque, Opaque, )
@@ -396,7 +396,7 @@ OPCODE(ImageSampleDrefImplicitLod, F32, U32,
OPCODE(ImageSampleDrefExplicitLod, F32, U32, Opaque, F32, Opaque, Opaque, )
OPCODE(ImageGather, F32x4, U32, Opaque, Opaque, Opaque, )
OPCODE(ImageGatherDref, F32x4, U32, Opaque, Opaque, Opaque, F32, )
-OPCODE(ImageFetch, F32x4, U32, Opaque, U32, U32, )
+OPCODE(ImageFetch, F32x4, U32, Opaque, Opaque, Opaque, Opaque, )
OPCODE(ImageQueryDimensions, U32x4, U32, U32, )
// Warp operations
diff --git a/src/shader_recompiler/frontend/maxwell/maxwell.inc b/src/shader_recompiler/frontend/maxwell/maxwell.inc
index d668dc1aa8..b47fb9c2e0 100644
--- a/src/shader_recompiler/frontend/maxwell/maxwell.inc
+++ b/src/shader_recompiler/frontend/maxwell/maxwell.inc
@@ -252,8 +252,8 @@ INST(SYNC, "SYNC", "1111 0000 1111 1---")
INST(TEX, "TEX", "1100 0--- ---- ----")
INST(TEX_b, "TEX (b)", "1101 1110 10-- ----")
INST(TEXS, "TEXS", "1101 -00- ---- ----")
-INST(TLD, "TLD", "1101 1100 --11 1---")
-INST(TLD_b, "TLD (b)", "1101 1101 --11 1---")
+INST(TLD, "TLD", "1101 1100 ---- ----")
+INST(TLD_b, "TLD (b)", "1101 1101 ---- ----")
INST(TLD4, "TLD4", "1100 10-- ---- ----")
INST(TLD4_b, "TLD4 (b)", "1101 1110 11-- ----")
INST(TLD4S, "TLD4S", "1101 1111 -0-- ----")
diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp
index 6a580f8319..60d61ec6eb 100644
--- a/src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/not_implemented.cpp
@@ -313,14 +313,6 @@ void TranslatorVisitor::SYNC(u64) {
ThrowNotImplemented(Opcode::SYNC);
}
-void TranslatorVisitor::TLD(u64) {
- ThrowNotImplemented(Opcode::TLD);
-}
-
-void TranslatorVisitor::TLD_b(u64) {
- ThrowNotImplemented(Opcode::TLD_b);
-}
-
void TranslatorVisitor::TLDS(u64) {
ThrowNotImplemented(Opcode::TLDS);
}
diff --git a/src/shader_recompiler/frontend/maxwell/translate/impl/texture_load.cpp b/src/shader_recompiler/frontend/maxwell/translate/impl/texture_load.cpp
new file mode 100644
index 0000000000..b4063fa6e1
--- /dev/null
+++ b/src/shader_recompiler/frontend/maxwell/translate/impl/texture_load.cpp
@@ -0,0 +1,165 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <optional>
+
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "shader_recompiler/frontend/ir/modifiers.h"
+#include "shader_recompiler/frontend/maxwell/translate/impl/impl.h"
+
+namespace Shader::Maxwell {
+namespace {
+
+enum class TextureType : u64 {
+ _1D,
+ ARRAY_1D,
+ _2D,
+ ARRAY_2D,
+ _3D,
+ ARRAY_3D,
+ CUBE,
+ ARRAY_CUBE,
+};
+
+Shader::TextureType GetType(TextureType type, bool dc) {
+ switch (type) {
+ case TextureType::_1D:
+ return dc ? Shader::TextureType::Shadow1D : Shader::TextureType::Color1D;
+ case TextureType::ARRAY_1D:
+ return dc ? Shader::TextureType::ShadowArray1D : Shader::TextureType::ColorArray1D;
+ case TextureType::_2D:
+ return dc ? Shader::TextureType::Shadow2D : Shader::TextureType::Color2D;
+ case TextureType::ARRAY_2D:
+ return dc ? Shader::TextureType::ShadowArray2D : Shader::TextureType::ColorArray2D;
+ case TextureType::_3D:
+ return dc ? Shader::TextureType::Shadow3D : Shader::TextureType::Color3D;
+ case TextureType::ARRAY_3D:
+ throw NotImplementedException("3D array texture type");
+ case TextureType::CUBE:
+ return dc ? Shader::TextureType::ShadowCube : Shader::TextureType::ColorCube;
+ case TextureType::ARRAY_CUBE:
+ return dc ? Shader::TextureType::ShadowArrayCube : Shader::TextureType::ColorArrayCube;
+ }
+ throw NotImplementedException("Invalid texture type {}", type);
+}
+
+IR::Value MakeCoords(TranslatorVisitor& v, IR::Reg reg, TextureType type) {
+ const auto read_array{
+ [&]() -> IR::U32 { return v.ir.BitFieldExtract(v.X(reg), v.ir.Imm32(0), v.ir.Imm32(16)); }};
+ switch (type) {
+ case TextureType::_1D:
+ return v.X(reg);
+ case TextureType::ARRAY_1D:
+ return v.ir.CompositeConstruct(v.X(reg + 1), read_array());
+ case TextureType::_2D:
+ return v.ir.CompositeConstruct(v.X(reg), v.X(reg + 1));
+ case TextureType::ARRAY_2D:
+ return v.ir.CompositeConstruct(v.X(reg + 1), v.X(reg + 2), read_array());
+ case TextureType::_3D:
+ return v.ir.CompositeConstruct(v.X(reg), v.X(reg + 1), v.X(reg + 2));
+ case TextureType::ARRAY_3D:
+ throw NotImplementedException("3D array texture type");
+ case TextureType::CUBE:
+ return v.ir.CompositeConstruct(v.X(reg), v.X(reg + 1), v.X(reg + 2));
+ case TextureType::ARRAY_CUBE:
+ return v.ir.CompositeConstruct(v.X(reg + 1), v.X(reg + 2), v.X(reg + 3), read_array());
+ }
+ throw NotImplementedException("Invalid texture type {}", type);
+}
+
+IR::Value MakeOffset(TranslatorVisitor& v, IR::Reg& reg, TextureType type) {
+ const IR::U32 value{v.X(reg++)};
+ switch (type) {
+ case TextureType::_1D:
+ case TextureType::ARRAY_1D:
+ return v.ir.BitFieldExtract(value, v.ir.Imm32(0), v.ir.Imm32(4), true);
+ case TextureType::_2D:
+ case TextureType::ARRAY_2D:
+ return v.ir.CompositeConstruct(
+ v.ir.BitFieldExtract(value, v.ir.Imm32(0), v.ir.Imm32(4), true),
+ v.ir.BitFieldExtract(value, v.ir.Imm32(4), v.ir.Imm32(4), true));
+ case TextureType::_3D:
+ case TextureType::ARRAY_3D:
+ return v.ir.CompositeConstruct(
+ v.ir.BitFieldExtract(value, v.ir.Imm32(0), v.ir.Imm32(4), true),
+ v.ir.BitFieldExtract(value, v.ir.Imm32(4), v.ir.Imm32(4), true),
+ v.ir.BitFieldExtract(value, v.ir.Imm32(8), v.ir.Imm32(4), true));
+ case TextureType::CUBE:
+ case TextureType::ARRAY_CUBE:
+ throw NotImplementedException("Illegal offset on CUBE sample");
+ }
+ throw NotImplementedException("Invalid texture type {}", type);
+}
+
+void Impl(TranslatorVisitor& v, u64 insn, bool is_bindless) {
+ union {
+ u64 raw;
+ BitField<49, 1, u64> nodep;
+ BitField<55, 1, u64> lod;
+ BitField<50, 1, u64> multisample;
+ BitField<35, 1, u64> aoffi;
+ BitField<54, 1, u64> clamp;
+ BitField<51, 3, IR::Pred> sparse_pred;
+ BitField<0, 8, IR::Reg> dest_reg;
+ BitField<8, 8, IR::Reg> coord_reg;
+ BitField<20, 8, IR::Reg> meta_reg;
+ BitField<28, 3, TextureType> type;
+ BitField<31, 4, u64> mask;
+ BitField<36, 13, u64> cbuf_offset;
+ } const tld{insn};
+
+ const IR::Value coords{MakeCoords(v, tld.coord_reg, tld.type)};
+
+ IR::Reg meta_reg{tld.meta_reg};
+ IR::Value handle;
+ IR::Value offset;
+ IR::U32 lod;
+ IR::U32 multisample;
+ if (!is_bindless) {
+ handle = v.ir.Imm32(static_cast<u32>(tld.cbuf_offset.Value() * 4));
+ } else {
+ handle = v.X(meta_reg++);
+ }
+ if (tld.lod != 0) {
+ lod = v.X(meta_reg++);
+ }
+ if (tld.aoffi != 0) {
+ offset = MakeOffset(v, meta_reg, tld.type);
+ }
+ if (tld.multisample != 0) {
+ multisample = v.X(meta_reg++);
+ }
+ if (tld.clamp != 0) {
+ throw NotImplementedException("TLD.CL - CLAMP is not implmented");
+ }
+ IR::TextureInstInfo info{};
+ info.type.Assign(GetType(tld.type, false));
+ const IR::Value sample{[&]() -> IR::Value {
+ return v.ir.ImageFetch(handle, coords, offset, lod, multisample, info);
+ }()};
+
+ IR::Reg dest_reg{tld.dest_reg};
+ for (size_t element = 0; element < 4; ++element) {
+ if (((tld.mask >> element) & 1) == 0) {
+ continue;
+ }
+ v.F(dest_reg, IR::F32{v.ir.CompositeExtract(sample, element)});
+ ++dest_reg;
+ }
+ if (tld.sparse_pred != IR::Pred::PT) {
+ v.ir.SetPred(tld.sparse_pred, v.ir.LogicalNot(v.ir.GetSparseFromOp(sample)));
+ }
+}
+} // Anonymous namespace
+
+void TranslatorVisitor::TLD(u64 insn) {
+ Impl(*this, insn, false);
+}
+
+void TranslatorVisitor::TLD_b(u64 insn) {
+ Impl(*this, insn, true);
+}
+
+} // namespace Shader::Maxwell
--
cgit v1.2.3-70-g09d2